GPT-6.1 Terra system card: frontier model scores 97.10% on ExploitBench (dangerous-capability eval)

The OpenAI deployment-safety system card addendum for GPT-6.1 Terra reports an ACE-normalized mean capability score of 97.10% on ExploitBench, with a pass@1 score of 83.33%. These are benchmark metrics for autonomously finding and ranking exploitable vulnerabilities, published on OpenAI's deployment-safety site as part of the model's frontier-capability disclosure.

This is a dangerous-capability-eval disclosure: a frontier model now scores near-ceiling on an exploit-benchmark, meaning autonomous exploit discovery/prioritisation is approaching a practical, deployable capability — which compresses time-to-exploitation and raises the floor on patching cadence for anything internet-facing. It is the measurement behind the 'AI finds bugs faster than defenders patch' thesis and informs how defenders set vulnerability triage and fix windows.

Not a direct attack vector; the capability is autonomous vulnerability discovery/exploit prioritisation demonstrated by the model at benchmark level.

OpenAI GPT-6.1 Terra (frontier model) — capability-eval disclosure, not an affected product.

For defenders: factor machine-speed vulnerability discovery into patch prioritisation (auto-patch where possible, tighten fix windows); treat AI-generated exploit candidates as a threat to patching cadence.