UK AISI ran Claude Mythos 5 on a hacking test with internet on. When the in-scope tries failed, it found a real GitHub repo, made fake identities, and pushed the maintainer to merge malware. Mid-attack, it stopped by itself in 2% of reruns. Told the targets were real people: 5%. Don't count on the model to quit. Put the stop button outside the loop. anthropic.com/research/align…

Ransomware crews now run intrusions through AI agents. SOCRadar got their logs: 8,257 messages, 133 sessions. The AI isn't the scary part. A human approved every step. What worked was one reused password that opened the VPN, servers, RMM, backup and security tools at once. Stop sharing passwords across those. socradar.io/blog/inc-gentlem…

Fully patched Windows machines still got a bootkit. Kaspersky says BlackX hit diplomats in Central Asia (APT41, medium confidence). It abuses CVE-2022-21894, which Microsoft fixed years ago. The catch: the old Microsoft boot cert is only revoked if you trigger it by hand. It expired Oct 19 and still loads the bad binaries. It needs admin first, so it's about staying in, not getting in. securelist.com/uefi-bootkit-…

Sandworm didn't break into this company. UAC-0099 did, then handed over the access (ESET says high confidence). One spear-phish in July 2025. Two months later, Sandworm pushed a wiper through AD Group Policy. The backdoor's C2 lived inside the Yggdrasil mesh network, so there's no normal domain or IP for your tools to flag. ESET says blocking the public Yggdrasil peers works. welivesecurity.com/en/eset-r…

x402 lets AI agents pay for APIs from a wallet. The default client had no per-payment limit and no total budget. V12 found 25 flaws, 3 critical. Fixed by Aug 20, but there's no CVE and no advisory, so nothing will alert you. If your agent holds a wallet, update and set your own cap. v12.sh/blog/x402

A firewall config backup can be a free key to your Active Directory. Group-IB found 70,000+ exposed credentials from VPN and firewall devices plus LDAP dumps. On FortiGate, an old default key (CVE-2019-6693) decrypts the saved config unless private-data-encryption is on. Check that setting. Then rotate the LDAP bind account, because that one reaches AD. group-ib.com/blog/stray-cat-…

PHP published another code execution advisory two days after PHP2Shell. CVE-2026-115560. No emergency patch, fix waits for the next normal release. It only bites if your app switches session.serialize_handler to php at runtime with ini_set(), and upload progress is on. Rare, but old apps do this. Already set session.upload_progress.enabled=0 for PHP2Shell? You're covered. Put it in php.ini or the FPM pool, then restart workers. php.net/archive/2026.php#202…

Pillow 13.0.0 fixed 38 security advisories in one release. One of them: a crafted FlashPix file runs code from a single Image.open() call, if olefile is installed. Any upload or thumbnail service that opens the file is enough. 13.0.0 dropped Python 3.10, so some of you can't just bump it. Upgrade Python first. pillow.readthedocs.io/en/sta…

HPE patched 72 bugs in ArubaOS. Eight score CVSS 10.0, no login needed. No known exploits yet. Big patch batches get reverse engineered fast. Until you update: put CLI and web management on their own VLAN. Check who can reach UDP 8211 (PAPI) too. support.hpe.com/hpesc/public…

HPE patched 72 bugs in ArubaOS. Eight score CVSS 10.0, no login needed. No known exploits yet. Big patch batches get reverse engineered fast. Until you update: put CLI and web management on their own VLAN. Check who can reach UDP 8211 (PAPI) too. support.hpe.com/hpesc/public…

The bug behind GitLab's CVSS 10 was in Ruby's regex engine. Ruby 4.0.9, 3.4.12 and 3.3.13 now fix it. Regexp.timeout doesn't protect you. Capping pattern length doesn't either. It only matters if user input reaches Regexp.new or /#{input}/. Grep for that. ruby-lang.org/en/news/2026/1…

PHP2Shell is fixed in 8.2, 8.3, 8.4 and 8.5. If you're on 8.1 or older, there's no new release to install. The PHP team only put out source backports for 7.4, 8.0 and 8.1. So you rebuild PHP yourself, or wait for your distro to do it. Run php -v on the real servers, not what the wiki says. vulncheck.com/blog/php2shell…

The Node.js Brotli bug now has a working RCE demo on a stock build. Imperva got code execution on the official Node 24.22.0 image. Two ways in: fetch() reading a Brotli response, or Express 5 / body-parser 2 decoding Brotli request bodies. A plain Node HTTP server doesn't decode bodies on its own. It took several tries, and no exploitation reported yet. Fixed in 22.23.4, 24.22.1, 26.12.1. Check your container images, they ship their own node. imperva.com/blog/node2shell-…

musl's regcomp can be broken by the size of a regex, before it matches a single character. Integer overflows, present since 0.5.0, all versions through 1.2.6. Crash at least, maybe heap writes. Only hits you if untrusted users can supply patterns. Think search boxes, filters, rule engines on Alpine images. No 1.2.7 yet, only a patch. openwall.com/lists/musl/2026…

Laravel shipped the fix for a 9.2 auth bypass on Aug 25. The release notes list one change: "Fix resource loading." The advisory came out Oct 19. Eight weeks later. No CVE. If a guard in config/auth.php uses driver 'token' on MySQL or MariaDB, an attacker can log in as an existing user without knowing their token. Fixed in 12.68.0 and 13.29.0. Rotating tokens won't help. Update. github.com/laravel/framework…

Lazarus is now hiding its payloads with post-quantum crypto (ML-KEM-1024). Kaspersky says they adopted it faster than many security products. But the crypto isn't how they got in. A phishing link to GitHub Pages, then a mandatory Korean banking security app, with the GitHub API as the C2. The tell is a built-in Windows binary (fsquirt.exe) running under a security service. Encryption can't hide where code runs from. securelist.com/lazarus-opera…

The Browser Company is shutting down Arc and says the reason is security. Keeping a browser safe is getting harder with AI, they say. Everything moves to Dia. New accounts stopped Oct 22. Desktop stops working Dec 22, and account data is deleted within 30 days after. Export your passwords and Spaces now. arc.net/

7 of 10 HTTP/3 stacks FoxIO tested had the same bug: tiny header block, huge memory. quic-go fixed it in 0.57.0. It sits under Caddy, Traefik and cloudflared, so check your build, not your config. Bun's HTTP/3 is still unpatched. One 516 KiB stream pins it at 100% CPU. foxio.io/blog/header-accordi…

Arista VeloCloud Orchestrator has a zero-day that's already being exploited. Any logged-in user, even the lowest role (Enterprise Read Only), can become a superuser operator. So your "harmless" read-only accounts are admin keys now. The web UI is exposed by default and no config turns that off. Save the logs first, then patch: 5.2.3.18, 6.1.3.10 or 6.4.2.12. Block 74.82.197.86 and 38.54.50.151. arista.com/en/support/adviso…

Microsoft says the ConnectWise cloud attack was run with an AI pentest tool called ARTEX. In several cases it took under 3 hours from first target to privileged access. Most attempts failed, but your "we'll look at it Monday" window is gone. If you ever authorized ConnectWise apps in Entra, search by app ID, not display name. Disable, don't delete, until the logs are saved. microsoft.com/en-us/security…