GPT-6 Sol · October Chat: how it does on security work, and what it will answer

What happened

OpenAI’s tuned October Chat version of GPT-6 Sol is evaluated separately from the earlier Work and Codex models. The vendor reports 82.62% on ExploitBench and 68.85% on SEC-Bench Pro at maximum reasoning effort. OpenAI classifies both October Chat models as High in cybersecurity, below its Critical threshold.

ExploitBench, score: 82.62% (vendor-card; OpenAI October GPT-6 safety card, maximum reasoning effort; ExploitBench as run by OpenAI; safeguards: Evaluation safeguard configuration is not specified for this score; run date not published; CONFIRMED). The benchmark includes historical public vulnerabilities, creating possible prior-exposure effects. The safety card was published on 7 October 2026.. SEC-Bench Pro, score: 68.85% (vendor-card; OpenAI October GPT-6 safety card, maximum reasoning effort; SEC-Bench Pro as run by OpenAI; safeguards: Evaluation safeguard configuration is not specified for this score; run date not published; CONFIRMED). Vendor evaluation; compare like-for-like harnesses and model versions. The safety card was published on 7 October 2026.. Refusal gate (CONFIRMED, observed 2026-10-08): Chat safeguards apply; the October rollout leaves Work and Codex models unchanged.. The Critical classification is a threshold in OpenAI’s Preparedness Framework. It is separate from an independent security assessment or a model’s availability in other products.