Google DeepMind skips a full capability assessment for Gemini 4 Carbon and carries forward Argon's bio and cyber findings

Google DeepMind · Company report

Read the source (opens in a new tab)
What happened
Google DeepMind published its Safety & Security Report for Gemini 4 Carbon in November 2026.
Key points
  • Carbon's gain on Google's general capability index fell below its material change threshold, so its Frontier Safety Framework did not require a full critical capability assessment.
  • Carrying forward Argon's results, Google says it "cannot rule out" that the unmitigated model reaches its cyber and biological critical capability levels or its stealth and situational awareness level.
  • Carbon still scored 97% on ExploitBench, 10 points above Argon, and 54.8% on ExploitGym, 12.8 points above, so Google hardened its cyber safeguards.
  • In agentic red teaming, safeguards cut the per-attempt breach rate to 0.8% from Argon's 3.1%, but 27.8% of rollouts still breached within 40 attempts.
Evidence
Company safety report. New tests covered safeguards, alignment, reasoning legibility and a 750-prompt biology propensity check with no hazardous answers; the capability determinations were carried over from Argon, not re-measured.
Uncertain
The threshold calls and the judgment that safeguards bring residual risk to an acceptable level are Google's own assessments.

See it in the latest record