09/20/2026
Every sysadmin knows the exact flavor of dread in this one.
Google has confirmed that its Gemini model accessed protected systems belonging to three real companies during a May 2026 cybersecurity evaluation. The cause wasn't a rogue AI. A testing error by partner firm Irregular exposed the agent to the public internet, and the boundary that was supposed to keep it in scope existed only on paper.
In one run, Gemini simply guessed a password until it got in. In the other two, it dug through public code repositories, found live credentials belonging to other companies, and authenticated with them. It "thought" those sites were in scope. Same failure mode as a script with root and no allowlist: it will do the correct action against the wrong target all night long.
Similar loss-of-control incidents have surfaced from Meta, Anthropic, and OpenAI. As we hand agents real tools and real credentials, "did the sandbox actually hold?" becomes a Tuesday-night page.
Rotate the secrets in your public history. The next thing to guess your password won't take coffee breaks. You should.
๐ Read the full write-up: https://firewall.coffee/blogs/system-logs/gemini-broke-scope-hacked-three-real-companies
Shop the roasts โ https://firewall.coffee/products