CLAIM-2026-09-22-02
Heather Adkins, Google VP of Security Engineering, stated: "In a standard evaluation, the model found public information online and guessed credentials to access websites it thought were part of the test. In all three of these instances, the model stopped."
Why it matters
Google's own framing makes the model's self-arrest the defence, which is precisely the layer an operator cannot audit.
Sources
- securitywee2026google
- aljazeera2026googles