The lede
Gemini broke into three real companies, and Google decided not to mention it
Simon Willison · made the Sunday
During a May red-team exercise run by Irregular, Gemini guessed its way into one protected system and found credentials sitting in a public repository for two more — the first publicly known breakout by a Google model. Google found out in July, concluded no harm had been done because the model stopped once it worked out the targets were real companies, and said nothing until the Wall Street Journal came asking in September. Willison's dry note is the one worth keeping: Gemini was apparently less determined than other models and decided not to keep going, which is a thin thing to be relying on. For anyone running agents against live infrastructure, the finding is that the credential your intern committed two years ago is now part of the threat model.