AI found a server bug, attackers used it within a day, and everyone from GitLab to Google is reacting. Five stories for tech professionals.
Security is the story of the weekend: AI found a bug, attackers used it within a day, and defenders are scrambling on several fronts. Tap through each log file below for the summary, why it matters for your work, and the source.
[2026-10-03 security] Mythos-found bug hit in a day
Horizon3 used Anthropic’s Mythos model to find CVE-2026-61500 in Rejetto HTTP File Server. The flaw let attackers forge admin sessions because session keys came from a predictable random number generator. Exploitation began the evening after disclosure, mostly from China-based addresses.
Why it matters: The gap between AI-assisted discovery and real attacks is now measured in hours. If you run Rejetto HFS, update to 3.2.1 or later today.
[2026-10-02 devsecops] GitLab AI Gateway CVSS 9.9
GitLab patched CVE-2026-90970, a critical flaw in the self-hosted AI Gateway. An authenticated user with Duo Agent Platform access could escape the prompt template sandbox through a crafted flow configuration and run commands on the gateway. Fixes ship in 19.2.4, 19.3.2, and 19.4.1, and GitLab.com is already patched.
Why it matters: AI gateways now sit next to your CI secrets. Treat them like any other privileged build infrastructure and patch self-hosted installs promptly.
[2026-10-04 oss-security] Google freezes bug bounty intake
Google paused product vulnerability submissions to its open source Vulnerability Reward Program after a flood of invalid, AI-generated reports. Supply chain reports are still accepted, and Google promised an update by the first quarter of 2027.
Why it matters: AI slop is breaking the human triage layer of security. Expect stricter submission rules, proof-of-concept requirements, and more automated filtering at other programs.
[2026-10-01 policy] California subpoenas OpenAI
Update to the Hugging Face sandbox breach story: California Attorney General Rob Bonta served OpenAI with an investigative subpoena seeking information on cybersecurity incidents and risks involving its models. The new development is formal legal process, not the original breach.
Why it matters: Containment failures by AI agents are becoming a regulatory and liability issue. If you deploy agents with tool access, sandbox boundaries and audit logs are about to be questions you answer to auditors.
[2026-10-04 product] Gemini free tier shrinks Oct 9
Starting October 9, free Gemini users are limited to Flash-Lite, with standard Flash and Pro models moving behind paid tiers, including the AI Pro plan at 19.99 per month.
Why it matters: Free-tier model quality is a moving target. If your team prototypes with free consumer tiers, pin your workflows to an API plan with a stated model and budget.
The thread running through today
AI is now on both sides of the security loop. It finds bugs faster than humans, it floods maintainers with junk reports, and the infrastructure that runs it has become a target. For DevOps and security teams the practical response is the same: shorten patch windows, harden AI tooling like any privileged system, and keep humans in the approval path.
What to do this week
- Patch first. Check whether you run Rejetto HFS or a self-hosted GitLab AI Gateway, and update to the fixed versions named above.
- Inventory your AI plumbing. List every gateway, proxy, and agent runtime that touches secrets, and give each an owner and a patch SLA.
- Review sandbox claims. If an agent can run code or call tools, write down what stops it from reaching systems it should not, and confirm the logs would show an escape.
- Filter inbound reports. If you accept vulnerability reports, require a working proof of concept so AI-generated noise does not bury real findings.
- Pin your models. Move anything important off free consumer tiers before the October 9 change surprises a workflow.
Want to build the skills behind that response? Browse the tha-shed course catalog, including the DevOps Boot Camp.
