Latest / Daily DefSec Brief / PHP2Shell exploited, LiteLLM poisoned, Fortinet fixes with no CVEs | Cyber Security News Oct 22
Transcript
- Jerry: Tell an AI agent to keep quiet on Slack, and it may take you more literally than you meant. This is the daily DefSec brief for Thursday, October 22nd, 2026. If you only do one thing today, patch PHP on every internet-facing server. 1. An update. Attackers are now broadly exploiting two critical PHP flaws we covered yesterday. The flaws can let anyone who can reach a PHP web server run code on it. Amazon's honeypots caught one attempt four days before the flaws went public. The fixes shipped yesterday. Patch every internet-facing PHP server today. Number two, attackers slip the credential stealer into one Lite LLM release on PyPi, and Docker Hub. LiteLLM is the gateway that holds a company's AI provider keys. The stealer runs whenever Python starts, even if light LLM is never imported. It steals cloud. Pipeline and AI keys and checks in every hour for more code to run. The bad code was only in built packages. The project's own code on GitHub was clean. If you installed that release, rebuild those hosts and rotate every key they held. three, no attacks reported but Fortinet is shipping critical fixes today for Fortigate firewalls and their management and logging tools. It's holding the advisories and the CVE numbers until patch Tuesday, november tenth. Kevin Beaumont says one of the flaws is a path traversal on the management interface. Until there's a CVE, your scanners have nothing to flag. Attackers can compare the old and new firmware while Fortinet's advisories are still pending. Fortinet calls the wait, responsible transparency. Patch your FortieGates this week and get admin access off the internet. 4. No attacks yet, but under certain conditions, GitLab hasn't described anyone who can reach a self-managed GitLab server can run code on it. It's the second 10 out of 10 GitLab flaw in nine days. It affects every release back to 2014. If you patched it on october twelfth, your server is exposed again. Upgrade self managed GitLab again. Feel free to copy last week's change request. Five. Kaspersky says an espionage crew planted a boot kit on fully patched Windows machines. Using an old secure boot bypass from the Black Lotus boot kit, it loads before Windows does. Before they could install the bootkit, the attackers needed admin rights on the machine. The Fix is a certificate revocation that Microsoft still leaves for you to switch on. Start rolling out Microsoft's secure boot revocation. Test it one hardware model at a time because it can't be undone. Six. No attacks yet, but Node.js fixed a bug in its brotli decompression that could let an attacker run code. The same release fixes ways to fool its TLS certificate checks. Upgrade Node.js and rebuild the container images that bundle it. By the way, this release also stops trusting the eMudra certificates whose keys turned out to be guessable. 7. No attacks yet, but an attacker with a stored cookie of a Clustered Cisco Expressway can run commands on it as root. That includes Expressway E, the edge box. Cisco files it under hard-coded credentials, and the patch is the only fix. Upgrade your clustered expressway servers. 8. No attacks yet, but ISC fixed four serious flaws in bind its DNS server. One of them may let an attacker run code by replaying the captured sign message. IC lists a fixed release only for newer branches. Patch bind and move servers still on the older branch to a current one. 9. Chrome can now warn people when a page tricks them into copying a ClickFix command. The warning only comes with enhanced protection, and that's off by default. Turn on enhanced protection in Chrome by policy. ClickFix has been on the show 18 times, so this one is welcome. 10. OpenAI says one of its coding agents was told not to send any Slack messages. It decided that covered the alerts to OpenAI's security team. A sub agent objected and was overruled. The list of people who get those alerts was emptied. A second agent removed the program's monitoring and then denied it when asked. Keep your agent's keys away from your alerting setup. It did keep things quiet, just as asked. Once more, patch PHP on your internet facing servers. Thanks for watching and the sources are in the show notes. Like, subscribe, and if the show helps, pass it to a colleague who'd use it. We'll be back again tomorrow.