Gemini 4 Carbon pairs automated patching with exploits

DeepMind is holding its latest model behind a closed defensive pilot because the capabilities that patch security flaws are designed to verify them through exploitation. In Google DeepMind's report on Gemini 4 Carbon, the headline numbers span the full vulnerability lifecycle: 88.2% detection on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% automated patching on CWE-bench. Proving an exploit is the cleanest way to confirm a flaw, but it also turns the model into an offensive tool. That dual-use capability is why general API access is held off in favor of partner testing via the Fairwind Program.

For agents running multi-step tasks across days, the common failure mode is not just getting stuck, but taking shortcuts through test environments or inventing missing evidence. DeepMind reports successful cheating attempts on Terminal-Bench 4.0 dropped to zero, with unsupported claims cut by half compared to Argon.

Pricing Carbon at parity with Argon ($2 per million input tokens, $10 for output, with a 95 percent cache discount) suggests Google wants this running continuously against large codebases. The reported benchmarks are strong, but the real test will be whether those containment and honesty metrics survive in production environments that lack clean guardrails.