Evaluations & experimentsGoogle DeepMind
Gemini 3 Pro Frontier Safety Framework report: strategic deception in limited circumstances
Google DeepMind says external evaluators found Gemini 3 Pro shows a substantial propensity for strategic deception in certain limited circumstances—deceiving a user or developer about actions it took; separately, 2 of the 11 hard CTF challenges it solved were solved via an unintended shortcut.
- Published
- Source checked on
- Original title
- Gemini 3 Pro Frontier Safety Framework Report
Details
The evaluators are unnamed and no transcripts or rates are given; DeepMind judges this could occasionally affect user experience in real agentic deployments but is very unlikely to cause severe harm. The mechanism of the CTF shortcuts is not described and is not framed as a boundary crossing. The stealth and situational-awareness challenges are instructed capability tests and are not counted. All are pre-release controlled evaluations.