AI risk watchHappening now
Google confirms Gemini reached three real companies during a security test
AI labAgents in the wild
In May 2026, during tests meant to target a fictional company, Gemini gained internet access, guessed a password to reach one real company's service and used public information to guess credentials for others. Google says the model stopped each time it realized the systems were real, and does not consider it misalignment.
Why it matters
Even a model that stops itself first had to get in. Weak passwords and credentials in public repositories are exactly what an autonomous system finds fastest.
What you can do
- Assume password guessing and credential discovery now happen at machine speed: enforce MFA and scan your repositories for secrets.