Monday, 26 October 2026

333 Tor exit nodes hijacked to intercept SSH traffic and retrieve plaintext email (opens in a new tab)

Ars Technica Security ·Dan Goodin ·fetched 26 Oct 2026, 07:40 UTC Must read agreed3/3

Why readMore than 300 Tor exit nodes were reconfigured to downgrade email to plaintext and funnel SSH through attacker-controlled hosts, with the Ebury Linux backdoor as the foothold.

An attacker reused a valid password to reach a Tor management server in late August, installed the Ebury backdoor, then pivoted over SSH to a second node called Switzerland2 using an RSA key investigators could not locate in its authorized_keys. From there they altered the exit policies of two hosted exit nodes to favour unencrypted mail traffic and redirect SSH connections. Over 333 exit nodes were affected, and the key mystery remains unexplained, which matters for anyone whose threat model assumed exit-node traffic was merely passively observed.

arstechnica.com → (opens in a new tab)

Leaked LiteLLM Master Key Let Attackers Steal GitHub Tokens and Hijack Release Pipeline (opens in a new tab)

The Hacker News ·The Hacker News ·fetched 26 Oct 2026, 03:42 UTC Must read agreed3/3

Why readAn unrevoked master key committed to a public repo in May let an attacker publish a malicious version of the LiteLLM Python package five months later.

A config file pushed to the public LiteLLM-Labs/lite-harness repository on 27 May 2026 contained the master key to a LiteLLM Agent Platform deployment; the commit was force-pushed over and the key flagged for rotation, but revocation never happened. On 10 October 2026 an attacker at 69.48.229[.]143 used the key to reach the Agent Platform, extracted a GitHub CLI OAuth token the service held for filing pull requests, and used it to push a malicious package release. Check whether you pulled LiteLLM around that window, and treat force-pushing over a leaked secret as what it is: not revocation.

Also covered byThe Record (opens in a new tab).

thehackernews.com → (opens in a new tab)

testmetest npm Packages Carry a Miasma-Family Worm (opens in a new tab)

SafeDep (supply chain) ·fetched 26 Oct 2026, 11:40 UTC Must read Research agreed3/3

Why readTracks a Miasma-family npm worm being tested in public, including the operator accounts, the 19-25 October publishing window, and why its loader layout defeats existing Miasma signatures.

An operator pushed 18 malicious npm packages between 19 and 25 October 2026 from accounts including alexjoipamiba and testnewtestme, testing a Miasma-family worm that steals credentials, spreads to other packages and repositories, and installs a wiper that deletes the home directory when the victim revokes the stolen GitHub token. The loader layout, infrastructure and marker strings differ from Miasma, defeating published signatures for loader size, file hashes and markers, and the token-revocation wiper is grafted from Mini Shai-Hulud. SafeDep found no compromised third-party packages or downstream victims as of 22 October; all activity traces to the operator's own accounts, and the 0.0.0-stage version stub is a shared npm artefact, not an indicator on its own.

safedep.io → (opens in a new tab)

PHP2Shell: Suspected North Korean campaign targets 216 vulnerable PHP hosts (opens in a new tab)

Security Affairs ·Pierluigi Paganini ·fetched 26 Oct 2026, 11:40 UTC CVE-2026-109619 EPSS 100.0% agreed3/3

Why readCyfirma names a suspected North Korean crew that has narrowed the global PHP2Shell exploitation wave down to a 216-host shortlist and coordinates over a private I2P channel.

Two critical PHP engine flaws patched on October 21, CVE-2026-109618 and CVE-2026-109619, are being hit at scale; Cloudflare counted millions of attempts within roughly a day of disclosure. Cyfirma reports that a suspected North Korea linked group is using AI to rank public exploits and has identified 216 hosts that still appear vulnerable, coordinating through an I2P channel the firm has monitored since October 22. The attribution rests on one vendor's assessment, so treat the DPRK link as provisional; the target count and the tradecraft are the parts worth acting on.

securityaffairs.com → (opens in a new tab)

Obstacles to Weaponizing PHP2Shell Likely Constrain Immediate Widespread Exploitation (opens in a new tab)

CrowdStrike ·Jaime Duque - Brenden Conrad ·fetched 26 Oct 2026, 03:42 UTC agreed3/3

Why readA read on how quickly PHP2Shell is likely to become commodity exploitation, which is what decides whether tonight is an emergency or a sprint.

CrowdStrike argues that practical obstacles to weaponizing PHP2Shell will constrain broad opportunistic exploitation in the near term, despite the attention the flaw is drawing. This is a vendor exploitability assessment rather than a measurement of observed activity, so it is useful for sequencing patch work and not for deciding to skip it. Read it alongside your own exposure count before letting the urgency drop, and note the vendor has a product interest in the framing.

crowdstrike.com → (opens in a new tab)

Ukraine warns hackers are exploiting Chinese-made Dahua surveillance cameras (opens in a new tab)

The Record ·fetched 26 Oct 2026, 15:42 UTC agreed2/2

Why readCERT-UA says attackers are abusing Dahua's Easy4IP cloud remote-access feature to take over internet-connected cameras and recorders in Ukraine.

CERT-UA identified cases where intruders changed device settings and created unauthorized user accounts on Dahua cameras and video recorders, exposing feeds and allowing remote control. The abuse centres on Easy4IP, the manufacturer's cloud service that connects devices using their serial numbers for remote access. The agency did not disclose device counts, timing or attribution; Dahua equipment remains in wide use by Ukrainian government institutions and military units.

therecord.media → (opens in a new tab)

'Gheist' Weaponizes GitHub Tokens to Steal Cloud Secrets (opens in a new tab)

Dark Reading ·Elizabeth Montalbano ·fetched 26 Oct 2026, 19:39 UTC agreed3/3

Why readAn active supply-chain campaign is abusing GitHub tokens across thousands of developer repositories to pivot into cloud secrets, which is a prompt to audit and rotate PATs and Actions tokens now.

The campaign, tracked as Gheist, harvests credentials and sensitive data from developer repositories at scale and uses captured GitHub tokens to reach cloud secrets. The item as received is short on indicators and on the initial access vector, so pair it with the primary research when it lands. In the meantime the defensive actions are clear: inventory long-lived PATs, scope and expire Actions tokens, and check OIDC trust policies that let repository identities assume cloud roles.

Also covered byInfosecurity Magazine (opens in a new tab).

darkreading.com → (opens in a new tab)

uCreator Android Spyware Can Steal Telegram 2FA Passwords and Clone WeChat Accounts (opens in a new tab)

The Hacker News ·The Hacker News ·fetched 26 Oct 2026, 11:40 UTC agreed3/3

Why readLinks the rentable uCreator Android kit to Doctor Web's Baohuo backdoor through a shared namespace and C2, covering Telegram 2FA theft and WeChat account cloning.

uCreator hides inside working rebuilds of Telegram and WeChat distributed by sideloading or run inside the kit makers' own container app, letting operators lift Telegram two-factor passwords and clone WeChat accounts. Comparing Ovi's October 21, 2026 analysis against Doctor Web's October 2025 Baohuo report shows the same code namespace com.ucreator.tgjar, the same C2 server and matching command names; Doctor Web put the earlier infection count above 58,000 devices. Whether the kit authors or a customer ran that campaign is unresolved.

thehackernews.com → (opens in a new tab)

AI-Generated Fake Bank Calls and NFC App Enabled €380,000 Fraud, Dutch Police Say (opens in a new tab)

Security Affairs ·Pierluigi Paganini ·fetched 26 Oct 2026, 15:42 UTC agreed3/3

Why readA bank-helpdesk fraud combining AI-generated voice calls with a fake banking app that reads victims' cards over NFC, worth €380,000 from 170 victims in three weeks.

Rotterdam police arrested two men over a scheme in which AI-generated phone calls convinced victims their accounts were at risk, then persuaded them to install a fake banking app and scan their physical bank card with the phone. The app used NFC to copy card data, which accomplices then used to withdraw cash. Victims spanned the Netherlands and many were in their twenties and thirties, which undercuts the assumption that voice-based social engineering only lands on older targets.

securityaffairs.com → (opens in a new tab)

Serbian autocrat loses election following widespread reports of digital repression (opens in a new tab)

The Record ·fetched 26 Oct 2026, 19:39 UTC agreed2/3

Why readShows the political endgame of a documented mercenary surveillance campaign: two years of Cellebrite extractions and spyware infections on opposition devices, and then the leader behind them conceding.

Aleksandar Vucic admitted defeat in Serbia's prime ministerial race after a run of disclosures about state digital repression. Amnesty International documented Cellebrite surveillance software being used to extract data from the phones of opposition figures, protesters, prosecutors and journalists during sham interrogations and arrests, with spyware installed in some cases, and Citizen Lab published further findings in the final week. For anyone tracking commercial spyware, this is a rare case where the forensic record became an electoral liability rather than a footnote.

therecord.media → (opens in a new tab)

Lessons Learned From 'AdFection' and What To Do Next (opens in a new tab)

EFF Deeplinks ·Thorin Klosowski ·fetched 26 Oct 2026, 19:39 UTC agreed2/3

Why readWalks through why AdFection is structurally different from ordinary malvertising and lists the mitigations that actually narrow the targeting surface.

EFF unpacks the Citizen Lab AdFection campaign, in which an ad network was used to deliver zero-click spyware to journalists' devices with no interaction required beyond the ad loading. Neither malvertising nor zero-click delivery is new on its own; the novelty is using the behavioral ad-tech targeting stack as the selection mechanism, so the surveillance economy becomes the aiming system for the implant. This is secondary coverage of someone else's research, but the mitigation section and the argument about ad-tech as attack infrastructure are the parts worth your time.

eff.org → (opens in a new tab)

26th October – Threat Intelligence Report (opens in a new tab)

Check Point Research ·fetched 26 Oct 2026, 19:39 UTC agreed2/3

Why readThree incidents with concrete numbers attached, including a hosting provider that shut down customer environments after malware was planted on their servers.

Check Point's weekly bulletin reports that Japanese provider Onamae Desktop Cloud was compromised with malware installed on customer servers, potentially touching more than 26,000 hosts, prompting the provider to begin shutting environments down on October 21. It also covers 3,897 State Bank of India customer records surfacing online with identity documents and loan data included, and a breach of WordPress backup vendor TeamUpdraft that exposed management-service connection keys later abused against downstream sites. The TeamUpdraft item is the one with supply chain implications worth acting on, since those keys reach into customer WordPress installations.

research.checkpoint.com → (opens in a new tab)

PHP2Shell security flaws now sucessfully exploited in the wild (opens in a new tab)

BleepingComputer ·Sergiu Gatlan ·fetched 26 Oct 2026, 11:40 UTC Must read CVE-2026-109619 EPSS 100.0% agreed3/3

Why readFirst high-confidence confirmation of working PHP2Shell RCE in the wild against Apache mod_php running PHP 8.4.26, not just crashes.

Honeypots run by Adel Karimi captured a full exploitation chain for CVE-2026-109618 and CVE-2026-109619: an 800 KiB decoded cookie dump providing memory disclosure, a function address derived from that dump using the public exploit's offsets, and a shell command executing roughly 1.6 ms after the final byte of the execution request. Marcus Hutchins independently assessed it as the strongest in-the-wild evidence so far. With EPSS at 0.99997 and the flaw sitting in the PHP engine itself rather than any application, any internet-facing PHP host that has not taken the 21 October emergency patches should be treated as exposed.

bleepingcomputer.com → (opens in a new tab)

PHP2Shell: No login, no app code, no warning. What next? (opens in a new tab)

The Register Security ·fetched 26 Oct 2026, 19:39 UTC Must read CVE-2026-109619 EPSS 100.0% agreed3/3

Why readGives the exact fixed PHP versions (8.5.12, 8.4.27, 8.3.36, 8.2.35) for the two unauthenticated RCE bugs now being exploited, plus which SAPIs are in scope.

CVE-2026-109618 and CVE-2026-109619, nicknamed PHP2Shell, allow unauthenticated remote code execution from a single crafted web request and both carry CVSS v4.0 9.2. The bugs sit in PHP core rather than any framework, so PHP-FPM, mod_php, php-cgi, FrankenPHP and LiteSpeed are all affected, with mod_php rated highest risk by the PHP team. Details leaked ahead of embargo and fixes shipped within about a day; exploitation in the wild has been reported, which makes inventory, not patching, the hard part.

theregister.com → (opens in a new tab)

‘This Is Not a Drill’: Critical PHP Vulnerabilities Are Setting the Internet on Fire (opens in a new tab)

WIRED Security ·Matt Burgess, Lily Hay Newman ·fetched 26 Oct 2026, 11:40 UTC Must read agreed3/3

Why readCVE-2026-109618 and CVE-2026-109619 affect all PHP installations serving web requests back to version 5, with no mitigation other than updating.

The PHP Foundation shipped emergency fixes for two critical remote code execution flaws reaching every PHP install that serves web requests, going back to PHP 5 from July 2004. Only PHP 8.2 and newer get official patches, though some distributions and third-party vendors produce legacy fixes, which matters because old PHP persists unnoticed in long-lived web infrastructure. The Foundation's own wording leaves no room for workarounds: there are no complete mitigations, updating is the only safe choice.

wired.com → (opens in a new tab)

PHP releases emergency fixes for 2 critical vulnerabilities (opens in a new tab)

CyberScoop ·Greg Otto ·fetched 26 Oct 2026, 15:42 UTC Must read CVE-2026-109618 EPSS 100.0% agreed3/3

Why readTwo critical PHP RCE bugs, CVE-2026-109618 and CVE-2026-109619, both CVSS v4 9.2 and carrying an EPSS of 0.99997, with patches out and only reconnaissance seen so far.

The PHP project shipped emergency patches for two remote code execution flaws dubbed "php2shell", both rated 9.2 on CVSS v4.0. CISA says it has observed no exploitation as of publication, and researchers report activity limited to scanning and reconnaissance, which given the EPSS percentile is a window rather than a reprieve. Patch PHP across internet-facing estates now and watch for the shift from recon to exploitation.

Also covered byCERT-FR (ANSSI) (opens in a new tab).

cyberscoop.com → (opens in a new tab)

Hackers got a four-day head start on epic PHP2Shell vulnerability (opens in a new tab)

Ars Technica Security ·Tom Wilson and Rafe Rosner-Uddin, Financial Times ·fetched 26 Oct 2026, 07:40 UTC agreed3/3

Why readAmazon says the PHP2Shell flaws are being broadly exploited in the wild, and that it saw exploitation attempts four days before public disclosure.

AWS reported that it detected an attempt to exploit the PHP vulnerability four days ahead of coordinated disclosure, and that the flaws are now under broad exploitation. The bugs were found with AI tooling and sit in a language underpinning a very large share of internet-facing applications. Reported via the Financial Times with no indicators or affected-version detail attached, so treat the exploitation claim as the actionable fact and chase the vendor advisory for scope.

arstechnica.com → (opens in a new tab)

Hacker Exploits Critical PHP2Shell Flaws to Win $2,000 Bitcoin Challenge, Releases PoC (opens in a new tab)

Cybersecurity News ·Guru Baran ·fetched 26 Oct 2026, 07:40 UTC CVE-2026-109618 EPSS 100.0% agreed3/3

Why readA working remote code execution chain for two critical PHP bugs is now published on GitHub, and CVE-2026-109618 sits at an EPSS of 0.99997, essentially the ceiling.

A researcher using the handle killa chained CVE-2026-109618 and CVE-2026-109619, the pair known as PHP2Shell, to get command execution on a deliberately unpatched PHP server and collect a 0.02511679 BTC bounty, then released a proof of concept. Both carry CVSS v4.0 9.2: one corrupts memory during cleanup after an out of memory condition, the other is reachable only when requests are served through a real PHP page rather than the command line SAPI. Public exploit code plus a top percentile EPSS means internet facing PHP needs patching now, not on the next cycle.

cybersecuritynews.com → (opens in a new tab)

Hackers Weaponize Public PHP2Shell Exploit Code to Target Universities and Business Websites (opens in a new tab)

Cybersecurity News ·Tushar Subhra Dutta ·fetched 26 Oct 2026, 23:41 UTC CVE-2026-109619 EPSS 100.0% agreed3/3

Why readPublic exploit code is now being fired at live university and business sites, and the mod_php trigger path means your WAF may never see the request.

Working exploit scripts for CVE-2026-109618 and CVE-2026-109619 spread within days of the October 21 fixes and are being used against unpatched university and commercial web servers. Because the flaw sits in the PHP engine rather than in site code, clean application code offers no protection, and Apache deployments running PHP as a module are worst off since the bug triggers before any site code executes. Web application filters and firewalls may never inspect the malicious request, which leaves patching as the only reliable control.

cybersecuritynews.com → (opens in a new tab)

Microsoft releases emergency .NET security fixes for Google Brotli bug (opens in a new tab)

The Register Security ·fetched 26 Oct 2026, 19:39 UTC CVE-2026-111497 EPSS 36.0% agreed3/3

Why readMicrosoft shipped out-of-band .NET 8, 9, 10 and 11 updates for the Brotli buffer overflow CVE-2026-111497, with an exploitability assessment of More Likely and working exploits already public.

The Brotli flaw allows an unauthenticated attacker to read or corrupt process memory, giving information disclosure, denial of service or privilege escalation; Microsoft rates it 8.1 under CVSS v3.1 and urges updating as soon as possible. The same bug is driving patches across Node.js, Python and PHP, so the practical task is inventorying every runtime that statically bundles Brotli rather than patching one product. No in-the-wild exploitation is reported yet, but exploits exist and EPSS sits at 0.36, the 98th percentile.

theregister.com → (opens in a new tab)

CVE-2026-111497: Critical out-of-bounds write in the Node.js Brotli decoder (opens in a new tab)

Rapid7 ·Rapid7 ·fetched 26 Oct 2026, 19:39 UTC CVE-2026-111497 EPSS 36.0% agreed3/3

Why readOut-of-bounds write in the Brotli decoder bundled with Node.js 26.x, 24.x and 22.x, reachable wherever an application decompresses attacker-supplied data with node:zlib.

CVE-2026-111497, published by the Node.js project on 21 October 2026, lets attacker-controlled Brotli input corrupt process memory via the bundled decoder, with crash, process-memory disclosure or remote code execution as the stated outcomes. Node.js rates it High while Google as CNA assigns CVSSv4 9.2; exposure hinges on whether your code path feeds untrusted compressed data into node:zlib, which covers a lot of proxies, API gateways and upload handlers. As of 26 October there is no known in-the-wild exploitation and no KEV listing, but EPSS is 0.36 (98th percentile), so treat upgrade as near-term work.

Also covered byoss-security (opens in a new tab).

rapid7.com → (opens in a new tab)

Microsoft Releases Out-of-Band Security Updates for Brotli and .NET tarballs Vulnerability (opens in a new tab)

Cybersecurity News ·Abinaya ·fetched 26 Oct 2026, 23:41 UTC CVE-2026-111497 EPSS 36.0% agreed3/3

Why readOut-of-band .NET updates complete two October fixes, including a Brotli buffer overflow reachable through ASP.NET Core request decompression and HttpClient automatic decompression.

Microsoft shipped unscheduled .NET updates on 25 October 2026 covering .NET 8, 9, 10 and 11 RC3. CVE-2026-111497 is a buffer overflow in Brotli decompression that an unauthenticated attacker can use to read or corrupt process memory, exposed anywhere an app handles Brotli-compressed input via ASP.NET Core request decompression, HttpClient automatic response decompression or the System.IO.Compression Brotli APIs; a second issue covers unsafe ownership settings in the Linux distribution tarballs. EPSS is 0.36, in the 98th percentile, and teams that already applied the 13 October updates are still exposed.

cybersecuritynews.com → (opens in a new tab)

PHP Issues Second Windows Security Update in Five Days to Fix Brotli Code Execution Flaw (opens in a new tab)

The Hacker News ·The Hacker News ·fetched 26 Oct 2026, 19:39 UTC CVE-2026-111497 EPSS 36.0% agreed3/3

Why readPHP rebuilt its Windows binaries for 8.2 through 8.5 a second time in five days because the bundled Brotli 1.2.0 carries CVE-2026-111497, and php -v will not tell you whether you are patched.

New Windows packages 8.5.12-1, 8.4.27-1, 8.3.36-1 and 8.2.35-1 ship Brotli 1.2.1 and retain the October 21 PHP2Shell RCE fixes, since the PHP source is unchanged. The flaw is reachable through the cURL extension when automatic HTTP response decompression is enabled, which is off by default in cURL but commonly turned on by HTTP client libraries. Because the PHP version string does not move, check the published SBOM for the Brotli version to confirm which build you are running. EPSS is 0.36, in the 98th percentile.

thehackernews.com → (opens in a new tab)

Linux Flaws Exposed Before Fixes as Root Exploit Goes Public (opens in a new tab)

Infosecurity Magazine ·fetched 26 Oct 2026, 23:41 UTC agreed3/3

Why readA public PoC for ZDI-CAN-33367, an unpatched Linux kernel LPE in the networking stack, gives a root shell on AlmaLinux 8 with no CVE and no merged upstream fix.

ZDI disclosed six Linux kernel flaws to maintainers and public mailing lists on 20 and 21 October, per Sysdig Threat Research analysis published on 24 October. One, ZDI-CAN-33367, carries a working public exploit and a CVSS of 7.8: an unprivileged user can modify the in-memory contents of files they hold only read access to, through a code path the earlier Dirty Frag page-cache fix missed. TuxCare reproduced a root shell with the PoC on AlmaLinux 8, and several of the flaws still have no fix in any released kernel.

infosecurity-magazine.com → (opens in a new tab)

The Pwn Request Reckoning: What Still Works Against GitHub Actions (opens in a new tab)

Wiz ·Shay Berkovich ·fetched 26 Oct 2026, 15:42 UTC Must read Research agreed2/2

Why readUpdated threat model for pull_request_target after GitHub's 2026 hardening: which Pwn request paths the platform closed and which still work, with an audit checklist.

GitHub's checkout protections, read-only cache tokens and event policies make the default-case Pwn request exploit substantially harder, but they shift the vulnerability class rather than removing it. What still works now requires explicit opt-ins, outdated checkout action versions, unprotected triggers, or script injection into workflows that run with repository secrets. The post closes with a checklist of what to look for in your own workflows, and builds on earlier parts covering the base threat model and AI-powered actions.

wiz.io → (opens in a new tab)

EVSEFuzz : A Tool for Automated Testing of Protocol Implementations (opens in a new tab)

arXiv cs.CR (all) ·Soham Gumaste, Michael Khan, Roland Varriale, Rigel Gjomemo ·fetched 26 Oct 2026, 07:40 UTC Research agreed3/3

Why readA fuzzer-builder for OCPP and similar protocol implementations in EV charging infrastructure, with two CVEs already published from it.

EVSEFuzz derives message structures from Python datatype validation packages to generate both compliant and non-compliant OCPP traffic, so fuzzers can be built without hand-writing grammars for each implementation. It has been run against several open and closed-source EV supply equipment stacks and produced two assigned CVEs. The datatype-validation trick generalises beyond charging infrastructure to any protocol whose implementation ships schema definitions.

arxiv.org → (opens in a new tab)

Kernel eviction notice: What wesp.sys reveals about the Windows Endpoint Security Platform (opens in a new tab)

Elastic Security Labs ·Pavel Yosifovich ·fetched 26 Oct 2026, 19:39 UTC Must read Research agreed3/3

Why readReverse engineering of wesp.sys and espclient.dll, the undocumented binaries behind Microsoft's Windows Endpoint Security Platform, which is intended to move EDR agents out of the kernel.

Two WESP components already ship in Windows Insider builds: the kernel driver wesp.sys and the user-mode API espclient.dll. wesp.sys is a file system mini-filter registered at altitude 329500 that talks to user mode over a filter communication port, and although the platform is undocumented the binaries carry public symbols, so a kernel debugging session recovers much of the structure. This is Microsoft's answer to the July 2024 CrowdStrike update that bluescreened 8.5 million machines, replacing per-vendor kernel drivers with one it maintains, and it tells detection teams what their sensors will be able to see once the transition lands.

elastic.co → (opens in a new tab)

OpenSSL 4.1.0: What’s new, what’s removed, and what’s still broken (opens in a new tab)

Help Net Security ·Anamarija Pogorelec ·fetched 26 Oct 2026, 03:42 UTC agreed3/3

Why readOpenSSL 4.1.0 ships DTLS 1.3 (RFC 9147) with four known bugs in that same DTLS code, two of which can make an application misread connection state.

The release adds DTLS 1.3 support and the project documents four outstanding defects in the DTLS implementation it ships with, affecting anyone adopting DTLS 1.3 or the newly added DTLS support in the SSL listener API. Two of the bugs can cause an application to misread a connection's state, which is the kind of error that turns into an authentication or termination assumption you did not intend. Plan the upgrade, but do not treat the new DTLS paths as production-ready yet.

helpnetsecurity.com → (opens in a new tab)

Cloudflare Release Emergency WAF Ruleset Update Following Brotli Vulnerability (opens in a new tab)

Cybersecurity News ·Abinaya ·fetched 26 Oct 2026, 15:42 UTC CVE-2026-111497 EPSS 36.0% agreed2/2

Why readCloudflare's emergency WAF rules for the Brotli bug ship in Log mode only, and existing action overrides can turn them into immediate blocks on legitimate traffic.

The 25 October managed ruleset adds two rules tied to CVE-2026-111497, one matching request bodies carrying Content-Encoding: br and one covering the Next.js image optimization endpoint. Both are enabled in Log mode by default, so customers on defaults get monitoring rather than prevention and must review matches before enabling blocking. The Brotli rule fires on any declared Brotli-compressed body, including normal client traffic, so a match is not evidence of exploitation; it does not match Accept-Encoding: br.

cybersecuritynews.com → (opens in a new tab)

Witness: Open-source framework for verifying how your software was built (opens in a new tab)

Help Net Security ·Anamarija Pogorelec ·fetched 26 Oct 2026, 07:40 UTC agreed3/3

Why readA CNCF/in-toto build attestation tool you can actually pull and wire into a pipeline to prove which command and which materials produced a given binary.

Witness is a CLI that records build metadata, signs the resulting attestation and verifies it against policy before release. TestifySec built it and donated it to the in-toto project under the CNCF, where it is now community maintained. Useful for teams that need provenance beyond a signed artefact: it answers which machine built the file, what went in, and whether the chain matches a declared policy.

helpnetsecurity.com → (opens in a new tab)

Looking for a Language [Guest Diary], (Wed, Oct 21st) (opens in a new tab)

SANS ISC Diary ·fetched 26 Oct 2026, 11:40 UTC agreed3/3

Why readA concrete Kibana Discover workflow that pivots on HTTP Accept-Language headers in honeypot traffic to surface anomalous clients.

Working from roughly a million records in a DShield SIEM, the author uses Kibana Discover to carve out web traffic by embedded language preference in HTTP headers and treats unusual Accept-Language values as the starting point for hunting. The method transfers to any Elastic deployment that retains request headers, and the write-up shows the query path rather than just asserting the idea. It is a modest hunting heuristic, but a reproducible one.

isc.sans.edu → (opens in a new tab)

Work to protect DNS from quantum decryption attacks is speeding up (opens in a new tab)

The Register Security ·fetched 26 Oct 2026, 07:40 UTC agreed2/3

Why readICANN's CTO has moved the deadline for quantum-safe DNSSEC from about ten years to three to five, which compresses the planning window for anyone signing zones.

John Crain, ICANN's chief technology officer, says internet governance bodies are accelerating work to make DNSSEC resistant to quantum decryption, and that he now expects capable machines around the turn of the decade rather than a decade out. DNSSEC is awkward to migrate because it adapts PKI to DNS records, and the obvious fix of larger keys collides with response size limits and fragmentation. The practical takeaway is a scheduling one: zone operators and registry planners should treat post-quantum signing as a near-term roadmap item rather than a research topic.

theregister.com → (opens in a new tab)

Identity Amortization: Shared Resources and the Cost of Sybil Attacks (opens in a new tab)

arXiv cs.CR (all) ·Homayoun Maleki, Nekane Sainz, Jon Legarda, Igor Santos-Grueiro ·fetched 26 Oct 2026, 23:41 UTC Research agreed3/3

Why readShows how to check whether your Sybil defense actually charges attackers what you think it does when one resource can back several identities.

The authors formalise identity amortization, the gap that appears when a priced resource such as a phone number can underwrite multiple identities, and show the assumed per-identity cost can be overstated by as much as the sharing factor. They give exact tests, computable from prices and capacities alone, for whether the gap exists, where it starts, and how large it can get, plus a sharing cap that eliminates it. The practical read is for anyone calibrating an identity or reputation threshold against a resource an adversary can reuse.

arxiv.org → (opens in a new tab)

Quantum Cadence: Post-Quantum Consensus with Multiple Concurrent Proposers (opens in a new tab)

arXiv cs.CR (all) ·Mohammad Mussadiq Jalalzai, Tobias Klenze, Jovan Komatovic ·fetched 26 Oct 2026, 15:42 UTC Research agreed3/3

Why readShows a way around the hardest practical problem in post-quantum consensus: without BLS-style aggregation, quorum certificates become too large to broadcast, so this design removes certificates entirely.

Quantum Cadence replaces the slot consensus in Cadence with Quantum Chorus, a multi concurrent proposer protocol in which each validator counts quorums directly among the messages it receives rather than assembling and forwarding signed certificates. Safety and liveness do not depend on signatures at all; signatures are layered on only to provide accountability. The paper claims the optimal fast path latency is preserved, which is the usual cost of abandoning aggregation. Relevant to anyone planning a PQC migration for a system whose bandwidth budget assumes aggregate signatures.

arxiv.org → (opens in a new tab)

Introducing the Magnet AI Research Lab: Our first peer-reviewed study puts AI in digital forensics to the test (opens in a new tab)

Magnet Forensics ·Phil Froklage ·fetched 26 Oct 2026, 19:39 UTC Research agreed3/3

Why readA peer-reviewed study in the Journal of Forensic Sciences found an AI expert system interpreted media authentication results 22 percentage points more accurately than the human examiner average.

Magnet Forensics has launched an AI Research Lab and published its first study, which tested how accurately trained examiners and an expert system read media authentication output. The expert system beat the human average by 22 percentage points, a figure that matters for examiners who will be asked on the stand how they know a result is reliable. The framing is explicitly about admissibility and the research base behind forensic tooling rather than a product launch, though the lab is the vendor's own.

magnetforensics.com → (opens in a new tab)

Cross-request KV cache exposure in the GLM-5.3 Workers AI model, and our remediation with SGLang (opens in a new tab)

Cloudflare Blog ·Michelle Chen ·fetched 26 Oct 2026, 23:41 UTC Must read Research agreed3/3

Why readCross-tenant KV cache leakage in a hosted inference stack: 155 responses on @cf/zai-org/glm-5.3 returned other customers' data, with the SGLang root cause published alongside.

Between 27 August and 9 September 2026 a tenant isolation bug in the GLM-5.3 Workers AI model returned fragments of other customers' requests, affecting 155 responses, all to prompts longer than 681,280 tokens. Cloudflare was alerted by Gadi Evron of Knostic, pulled and restored the model, and found no evidence of deliberate exploitation, with request timing consistent with ordinary long-context use. The post is published as coordinated disclosure with SGLang, the open-source inference engine where the cache-reuse bug lives, so anyone self-hosting SGLang for multi-tenant serving should read the root cause rather than the incident summary.

blog.cloudflare.com → (opens in a new tab)

AI, Scripts, Operator: Inside the ConnectWise Attack (opens in a new tab)

Huntress ·fetched 26 Oct 2026, 19:39 UTC Must read agreed3/3

Why readFirst-party telemetry on what an AI-driven intrusion actually looked like on the wire, plus a blunt warning that the tells you would build detections on are temporary.

Huntress follows up its ConnectWise Cloud Backup credential abuse report with the telemetry behind it, after Microsoft attributed the activity to Storm-0940 using the ARTEX agent framework. On 15 October it observed three distinct modes: scripted enumeration, hands-on operator work, and tooling that kept firing at a target that was already gone. Huntress rates an LLM's involvement at low-to-medium confidence on its own data and medium-to-high with Microsoft's attribution, and argues that cadence, retry loops and user agent artefacts will disappear as agent tooling matures, so detections should target what the activity does rather than how it paces itself.

huntress.com → (opens in a new tab)

Trusting the Tool: Asymmetric Deference of LLM Vulnerability Auditors to Static-Analysis Evidence (opens in a new tab)

arXiv cs.CR (AI) ·ZhiHong Liu, Shuyao Gao ·fetched 26 Oct 2026, 07:40 UTC Must read Research agreed3/3

Why readMeasures that LLM vulnerability auditors fed fabricated static-analysis assurances miss up to 19 points more real bugs, while fabricated alarms barely hurt them.

The authors built an intraprocedural C/C++ analyser emitting proof obligations and ran 16,572 audits over 916 vulnerable/patched PrimeVul function pairs, across eleven model configurations from four vendors and six evidence conditions. True evidence improved accuracy by only 0.02, but false assurances that a missing check exists suppressed detection of real vulnerabilities by as much as 19 points, whereas false alarms were disputed and did not degrade accuracy. Anyone wiring a scanner into an LLM triage pipeline is building in an asymmetric trust failure that favours false negatives.

arxiv.org → (opens in a new tab)

AI coding agents fall for malicious instructions planted in the Downloads folder (opens in a new tab)

Help Net Security ·Sinisa Markovic ·fetched 26 Oct 2026, 11:40 UTC Must read Research agreed3/3

Why readA drive-by downloaded AGENTS.md in the Downloads folder makes a coding agent install a malicious npm package or hand over the user's login token.

Forever Security found that AGENTS.md, the plain-text instruction file coding agents read from a project directory, is honoured even when the directory is Downloads and the file arrived from a website without the user approving the save. Opening an agent in that folder is enough: the planted instructions got agents to install an attacker-chosen npm package and to disclose the user's token. The fix surface is the agent's trust in any AGENTS.md on disk, which means Downloads and other drive-by writable directories need treating as hostile input.

helpnetsecurity.com → (opens in a new tab)

AgentTrace: Execution-Level Fingerprinting of Black-Box LLM Agents (opens in a new tab)

arXiv cs.CR (AI) ·Gejian Zhao, Hanzhou Wu, Xinpeng Zhang ·fetched 26 Oct 2026, 11:40 UTC Research agreed3/3

Why readA method for identifying which foundation model sits behind a black-box LLM agent, reaching 96.0% attribution accuracy even when the agent actively spoofs its identity.

AgentTrace injects structurally ambiguous probes into an agent's rigid tool-call interfaces to trigger tokenization-sensitive serialization failures, then models the temporal sequence of the agent's repair attempts with lightweight Markov chains. Initial failures separate model families; the ordered recovery trajectory separates intra-family variants that share a tokenizer, such as Llama-3 8B versus 70B, which unordered behavioural statistics cannot do. The result is provenance verification that survives system-prompt obfuscation and identity spoofing, relevant to anyone needing to confirm what model an agent in their supply chain actually runs.

arxiv.org → (opens in a new tab)

TRACE: Data-Free Text Reconstruction Attacks against Approximate Unlearning in LLMs (opens in a new tab)

arXiv cs.CR (AI) ·Mengying Zhang, Derui Wang, Ibrahim Khalil, Bowen Liu ·fetched 26 Oct 2026, 23:41 UTC Research agreed3/3

Why readKeeping pre- and post-unlearning model snapshots for audit is itself a privacy side channel: TRACE reconstructs the supposedly removed training text from the parameter diff alone.

TRACE recovers unlearned training data from paired pre- and post-unlearning LLM snapshots with no data-specific prior knowledge. It narrows the search by extracting candidate tokens from embedding-level parameter differences, then assembles sequences using a contrastive score based on the perplexity gap between the two models, with a two-stage exploration-then-completion strategy. The practical consequence is that approximate unlearning plus snapshot retention for rollback or compliance audit can defeat the deletion it was meant to deliver.

arxiv.org → (opens in a new tab)

Detecting Insecure Compositions of Clean Agent Skills via Capability-Flow Graph Analysis (opens in a new tab)

arXiv cs.CR (AI) ·Yihao Li, Chuanpu Fu, Yizhou Zhao, Miao Chen ·fetched 26 Oct 2026, 07:40 UTC Research agreed3/3

Why readDemonstrates that individually clean agent skills chain into safety violations through the agent's own runtime handoffs, and gives a graph method to find those paths.

The risk is compositional: an upstream skill's output becomes a downstream skill's input, so no single skill is malicious and malicious-skill scanners see nothing. KnotScryer abstracts each skill into a capability summary, builds a Skill Capability-Flow Graph where edges are cross-skill information flows, and searches for paths carrying protected or untrusted data into insecure operations, with a two-layer verification step to cut false paths. Directly relevant to anyone running third-party skill or tool marketplaces on an agent platform.

arxiv.org → (opens in a new tab)

CVE-2026-57845 (CVSS 8.7): Hugging Face Datasets through 5.1.0 contains an insecure deserialization vulnerability that allows attackers to execute arbitrary Python code by craft (opens in a new tab)

NVD ·fetched 26 Oct 2026, 15:42 UTC CVE-2026-57845 CVSS 8.7 EPSS 0.5% agreed3/3

Why readNames the exact call that turns a shared dataset into code execution: load_dataset("pandas", data_files=...) hands untrusted files to pandas.read_pickle with no validation.

Hugging Face Datasets through 5.1.0 still ships the deprecated pandas packaged builder, whose Pandas._generate_tables path calls pandas.read_pickle() on the supplied data files. A pickle-backed dataset carrying a __reduce__ payload therefore executes attacker code with the privileges of whoever loads it, typically a researcher's workstation or a training job with credentials attached. Pickle executing code is expected; the finding is that a mainstream dataset loader routes untrusted remote files into it by default, so audit notebooks and pipelines for that builder and pin or remove it.

nvd.nist.gov → (opens in a new tab)

On Task, Off Track: Benchmarking LLM and Agent Robustness to Near-Target Prompt Injection (opens in a new tab)

arXiv cs.CR (AI) ·Yuyang Gong, Zihao Wang, Jiawei Liu, XiaoFeng Wang ·fetched 26 Oct 2026, 07:40 UTC Research agreed3/3

Why readIntroduces near-target prompt injection, attacks that preserve most of a legitimate answer while making it wrong or harmful, and a benchmark showing current defences miss them.

Existing injection benchmarks reward models for rejecting instructions that obviously abandon the user's task, which lets a system score well while still accepting subtle corruptions of a correct answer. NTBench builds target-instruction pairs from legitimate tasks in existing datasets, with separate tracks for incorrect and harmful outcomes in both responses and agent tool invocations, including harmful targets that accompany otherwise correct completion. Deviations are validated by a mix of LLM checks and human review, so the resulting scores reflect intended corruption rather than refusal behaviour.

arxiv.org → (opens in a new tab)

Bias in Qwen3-Coder Model (opens in a new tab)

Schneier on Security ·Bruce Schneier ·fetched 26 Oct 2026, 11:40 UTC agreed3/3

Why readQwen3-Coder writes measurably less secure code when told the customer is a US defence contractor or a Hong Kong democracy group, and the bias sits in the weights.

Hirundo's testing found Qwen3-Coder 19% more likely to return a hackable app when asked to build for a US defence contractor rather than a Chinese one, and 28% more likely to leak other users' data for a Hong Kong democracy group. Adding an explicit security instruction made the gap worse, not better: 27% versus a neutral customer. Removing the model's CCP alignment closed all three gaps while leaving coding ability on ten other tasks intact, which locates the behaviour in the weights rather than in a system prompt.

schneier.com → (opens in a new tab)

CVE-2026-115812 (CVSS 9.3): Relative path traversal during model archive extraction in Amazon Web Services Multi Model Server 1.1.11 and earlier might allow remote actors to exec (opens in a new tab)

NVD ·fetched 26 Oct 2026, 15:42 UTC CVE-2026-115812 CVSS 9.3 EPSS 0.5% agreed3/3

Why readRemote code execution from a crafted model archive in an AWS model server that is end of life, so migration is the only fix available.

AWS Multi Model Server 1.1.11 and earlier mishandles relative paths while extracting model archives, letting a crafted archive write outside the intended directory and reach arbitrary code execution. Amazon has declared the product end of life and will not ship a patched version, so anyone still serving models through it has to move to a supported serving stack rather than patch. Treat model archives as untrusted input in whatever you migrate to, because the same extraction pattern recurs across model packaging formats.

nvd.nist.gov → (opens in a new tab)

Label-Budgeted Attribute Inference from Protected Features in Edge-Cloud Split Multimodal LLMs (opens in a new tab)

arXiv cs.CR (AI) ·Junguang Yao, Peichun Hua, Hanxiu Zhang, Yue Zheng ·fetched 26 Oct 2026, 15:42 UTC Research agreed3/3

Why readShows that intermediate features sent from device to server in split multimodal LLM inference leak age, gender and race even when raw images never leave the device.

The authors propose a label-budgeted attribute-inference framework that measures representation-level privacy with a recovery curve: how much of the unprotected leakage an honest-but-curious server recovers as a function of its labelled-sample budget. Testing across two multimodal LLMs, unprotected split-inference pipelines leaked all three protected attributes from intermediate features. The method also exposes why existing defence evaluations, which rely on final-response refusals or a fixed classifier failing, understate what an adaptive server can do.

arxiv.org → (opens in a new tab)

Texas sues OpenAI for deceiving consumers about safety, seeks sweeping reforms and fines (opens in a new tab)

The Record ·fetched 26 Oct 2026, 23:41 UTC agreed3/3

Why readA state attorney general is testing whether an AI vendor's safety marketing is a deceptive trade practice, and whether the CEO is personally on the hook for it.

Texas has sued OpenAI and Sam Altman under state consumer protection law, arguing that ChatGPT is sold as safe while its guardrails fail, including for minors. The remedies sought go well past money: barring minors from the product, routing new model releases through a court-appointed expert, and identity verification the state says is already legally required. Naming the chief executive personally is the part to watch, because it sets a template other AGs can copy against any company making safety claims about a deployed model.

therecord.media → (opens in a new tab)

Developing Security Metrics for Automated Compliance Checking (opens in a new tab)

arXiv cs.CR (all) ·Immanuel Kunz, Nico Haas, Angelika Schneider, Christian Banse ·fetched 26 Oct 2026, 19:39 UTC Research agreed3/3

Why readProposes an ontology-backed data model and elicitation method for turning prose compliance requirements into measurable technical metrics, with an open-source implementation started.

Existing security metrics in the literature are inconsistent in data model and abstraction level, overlap heavily, and resist combination into full coverage of a standard's requirements. The authors systematise compliance requirements into a data model and define a methodology for deriving metrics at a comparable abstraction level, plus early work on a collaborative open-source metric catalogue. Relevant if you are building continuous compliance or audit automation against ISO 27001, C5 or similar rather than hand-mapping controls.

arxiv.org → (opens in a new tab)

Google Says Gmail and Photos Data Shared With AI Search Could Be Used for Targeted Ads (opens in a new tab)

The Hacker News ·The Hacker News ·fetched 26 Oct 2026, 07:40 UTC agreed3/3

Why readGoogle's privacy policy, effective 16 October 2026, now permits Search and Gemini to read Gmail, Drive and Photos data and use it for ad personalisation when a user opts in.

The revised policy states that users may allow Search or Gemini Apps to access data from other Google services, and that Search may process that data to personalise content including ads. Google maintains it does not serve personalised ads from Drive, Gmail or Photos content unless the user enables cross-service access. For enterprise Workspace tenants this is a data-governance question worth checking against admin controls and DPIA assumptions before users flip the setting themselves.

thehackernews.com → (opens in a new tab)

CISA 2030 Sets Out ‘Aggressive, Anticipatory’ Cyber Defense Strategy for Critical Infrastructure (opens in a new tab)

Infosecurity Magazine ·fetched 26 Oct 2026, 19:39 UTC agreed3/3

Why readCISA published its CISA 2030: Defend, Harden, Prevail strategy on October 26, signalling a shift from passive defence to a proactive posture on critical infrastructure.

Acting director Nick Anderson frames the strategy around faster threat response and building resilient systems rather than reacting after the fact. The document sets direction for how CISA will engage critical infrastructure operators through 2030. It is a statement of intent rather than a rule, so it changes no obligation today, but it is the frame future CISA requirements and engagement will be built on.

infosecurity-magazine.com → (opens in a new tab)

UNM Health cyberattack disrupts care statewide (opens in a new tab)

Google News: incidents · Albuquerque Journal ·fetched 26 Oct 2026, 23:41 UTC agreed3/3

Why readA cyberattack at UNM Health is disrupting patient care across New Mexico, the kind of healthcare incident that lands on a hospital board's Monday agenda.

UNM Health is reporting statewide disruption to care following a cyberattack. The item arrived as a headline only, with no detail on the attack vector, the actor or the systems affected. The signal for a security leader is the operational impact on a state health system rather than anything technical to act on yet.

Also covered byKRQE (opens in a new tab).

news.google.com → (opens in a new tab)

Disgraced eMudhra drops out of TLS business after using cert private key of 4 (opens in a new tab)

Ars Technica Security ·Dan Goodin ·fetched 26 Oct 2026, 07:40 UTC agreed3/3

Why readeMudhra is exiting the public TLS business after issuing eight intermediates with keys weak enough to mint certificates for any site, one with a private key of 4.

Chairman V. Srinivasan said the India-based CA will withdraw from global SSL/TLS, citing compliance cost, free certificate issuance eroding returns, and TLS contributing roughly 1 percent of revenue, while complaining that social media amplification brought disproportionate scrutiny. The underlying failure was eight intermediate credentials so weak that anyone could issue certificates for any website or mail server. For anyone tracking CA trust decisions or chains terminating at eMudhra, this is the end state of that incident.

arstechnica.com → (opens in a new tab)

Japan Government to Hold Ministerial Meeting on Cyberattack Response Tonight; PM Takaichi to Order Emergency Measures (opens in a new tab)

Google News: incidents · BigGo Finance ·fetched 26 Oct 2026, 07:40 UTC agreed3/3

Why readJapan's cabinet is convening at ministerial level over an ongoing cyberattack, with PM Takaichi expected to order emergency measures, which sets the escalation bar for peers operating in the region.

The Japanese government is holding a ministerial meeting on cyberattack response, with the prime minister expected to direct emergency measures. The report gives no detail on affected entities, actor or vector. National-level intervention of this kind signals an incident with sector-wide operational impact, and anyone with Japanese operations or suppliers should expect questions about exposure.

news.google.com → (opens in a new tab)

Korean Air and Asiana Airlines Data Breach via Amadeus: Cybersecurity Incident Analysis and Third-Party Risks (opens in a new tab)

Google News: incidents · Rescana ·fetched 26 Oct 2026, 07:40 UTC agreed3/3

Why readKorean Air and Asiana passenger data exposed through Amadeus, the shared GDS platform most of the airline sector depends on.

Both carriers are reported breached via their common third party, Amadeus, rather than through their own systems. The exposure is a single-supplier concentration risk: any airline using the same global distribution system is in the same position, which is the question a board in that sector will ask this week. Detail in the available text is thin, so treat scope and record counts as unconfirmed pending carrier statements.

news.google.com → (opens in a new tab)

23andMe successor agrees to $6M settlement with California over data breach (opens in a new tab)

Google News: incidents · Top Class Actions ·fetched 26 Oct 2026, 15:42 UTC agreed3/3

Why readA $6M California settlement over the 23andMe genetic-data breach gives a concrete regulator price tag for consumer-genomics data loss, now carried by the post-bankruptcy successor entity.

The successor to 23andMe has agreed to pay California $6 million to resolve state claims arising from the credential-stuffing breach that exposed customer genetic and relative-matching data. The settlement lands after 23andMe's bankruptcy and asset sale, so the liability followed the data into the acquiring entity. Only the headline figure and the parties are reported here; the feed text carries no settlement terms or injunctive obligations.

news.google.com → (opens in a new tab)

Japan companies compensate users as cyberattacks buffet economy (opens in a new tab)

Google News: incidents · Reuters ·fetched 26 Oct 2026, 03:42 UTC agreed3/3

Why readReuters reports Japanese firms paying compensation to customers as the cumulative cost of cyberattacks starts showing up at economy scale.

Companies in Japan are compensating users following a run of cyberattacks, a sector-level signal that incident cost is moving from remediation budgets into direct customer redress. Only the headline reached us, so treat the detail as unverified, but the direction is the thing a board will ask about: what our equivalent exposure and payout policy looks like.

news.google.com → (opens in a new tab)

Britain bets big on AI to defend against the threats AI is helping create (opens in a new tab)

The Register Security ·fetched 26 Oct 2026, 11:40 UTC agreed3/3

Why readWhat the UK is actually buying with £130 million of cyber money, including a regulatory sandbox for autonomous security agents on critical infrastructure.

Ahead of this week's Budget, the Treasury earmarked £130 million inside a wider £600 million-plus AI package: £60 million for a National Cyber Shield built on agentic AI, covering a £25 million challenge fund for UK security firms and a sandbox to test autonomous tooling across critical infrastructure, government and industry, plus £70 million for a National Exposure and Real-Time Vulnerability Centre. The announcement lands days after PHP2Shell knocked over government websites. Nothing is deployed yet, so the usable signal is the funding direction and the sandbox, which is where procurement and assurance questions will eventually surface.

theregister.com → (opens in a new tab)

UK Biobank set to reopen to researchers after data breach (opens in a new tab)

Google News: incidents · Research Professional News ·fetched 26 Oct 2026, 07:40 UTC agreed3/3

Why readUK Biobank is restoring researcher access after a breach, a named incident in a sector that holds genomic and health data at scale.

UK Biobank is preparing to reopen its data platform to researchers following a data breach that suspended access. The report carries no detail on intrusion vector, scope of exposed records or timeline. Worth tracking for anyone running research data platforms or answering questions about custodianship of sensitive cohort data.

news.google.com → (opens in a new tab)

Federal Courts Review Alleged Transcript Data Breach (opens in a new tab)

Google News: incidents · Law360 ·fetched 26 Oct 2026, 11:40 UTC agreed3/3

Why readThe US federal judiciary is reviewing an alleged breach of court transcript data, a disclosure a legal-sector board will field on Monday.

Federal courts are reviewing an alleged data breach involving court transcript data. The report names the judiciary as the affected body but carries no technical detail on scope, method or exposure. It registers as an incident disclosure at a significant public institution rather than as a campaign writeup.

news.google.com → (opens in a new tab)

Alabama Power parent company sued over data breach (opens in a new tab)

Google News: incidents · AL.com ·fetched 26 Oct 2026, 15:42 UTC agreed3/3

Why readLitigation has followed the breach at Alabama Power's parent, the kind of post-incident legal exposure a board will ask about.

AL.com reports that the parent company of Alabama Power is being sued over a data breach. The item arrived as a headline with no detail on the claims, class size or the underlying incident. Relevant mainly as a marker of civil liability following utility-sector breach disclosure.

news.google.com → (opens in a new tab)

GIANT Back Online After Cyberattack, Shopper Discount Ends Thursday (opens in a new tab)

Google News: incidents · PHILADELPHIA.Today ·fetched 26 Oct 2026, 11:40 UTC agreed3/3

Why readGIANT has restored operations after a cyberattack, with a customer discount closing Thursday.

The regional grocery chain is back online following a cyberattack, and the goodwill shopper discount offered during the outage ends Thursday. The item arrived as a headline with no body, so cause, duration and data impact are not covered. Useful only as a recovery datapoint for retail peers.

news.google.com → (opens in a new tab)

Keystone data breach: customer addresses exposed (opens in a new tab)

Google News: incidents · CryptoTicker ·fetched 26 Oct 2026, 03:42 UTC agreed3/3

Why readKeystone discloses a breach exposing customer postal addresses, which for a hardware wallet vendor is a physical-safety problem as much as a privacy one.

Customer addresses were exposed in a breach at Keystone, reported by CryptoTicker. Only the headline reached us, so scope, cause and record count are unknown. Address leaks from crypto hardware vendors have historically preceded targeted extortion and in-person attacks, which is the part worth tracking as detail emerges.

Also covered bybloomingbit (opens in a new tab).

news.google.com → (opens in a new tab)