Gemini 4 Carbon: how it does on security work, and what it will answer

What happened

Google reports 88.2% on SEC-Bench Pro, 54.8% on ExploitGym and 71.0% on CWE-bench for Carbon. These launch evaluations ran with guardrails disabled. Initial product access is for selected cyber defenders through Fairwind.

SEC-Bench Pro, vendor-reported score: 88.2% (vendor-card; Google DeepMind launch evaluation; detailed configuration not specified; SEC-Bench Pro as evaluated by Google DeepMind; safeguards: Guardrails disabled for this evaluation; run date not published; CONFIRMED). Vendor-reported percentage. Run date, task count, version and budget are unspecified; this is separate from deployed-product safeguards.. ExploitGym, vendor-reported score: 54.8% (vendor-card; Google DeepMind launch evaluation; detailed configuration not specified; ExploitGym as evaluated by Google DeepMind; safeguards: Guardrails disabled for this evaluation; run date not published; CONFIRMED). Vendor-reported percentage. Run date, task count, version and budget are unspecified; this is separate from deployed-product safeguards.. CWE-bench, vendor-reported score: 71% (vendor-card; Google DeepMind launch evaluation; detailed configuration not specified; CWE-bench as evaluated by Google DeepMind; safeguards: Guardrails disabled for this evaluation; run date not published; CONFIRMED). Vendor-reported percentage. Run date, task count, version and budget are unspecified; this is separate from deployed-product safeguards.. Refusal gate (CONFIRMED, observed 2026-10-28): Initial access for selected cyber defenders through Fairwind; wider access planned after further feedback and safeguards work.. The announcement date identifies the source publication. Evaluation run dates, task counts, versions and detailed budgets are unspecified. These vendor-reported percentages should be compared with runs using the same conditions. The evaluation configuration is separate from safeguards and access controls in the planned product rollout.