GPT-6.1 Sol: how it does on security work, and what it will answer

What happened

OpenAI released GPT-6.1 Sol on 29 September 2026 and treats it as Critical in cybersecurity, with the same safeguards as GPT-6 Astra. Its own addendum gives 99.7% on ExploitBench, with a contamination caveat, and 21.5% arbitrary code execution on an internal port of vulnerabilities disclosed from June to August 2026, against 31.5% for GPT-6 Astra. It scores 78.8% pass@1 on SEC-Bench Pro and 35.1% on ExploitGym.

ExploitBench, score: 99.7% (vendor-card; OpenAI GPT-6.1 Sol addendum to the GPT-6 Astra system card, 29 September 2026; ExploitBench as run by OpenAI; safeguards: safeguard state during the run unstated; OpenAI ran its evaluations in its research environment or through the API; run 2026-09-29; CONFIRMED). Maximum reasoning effort; OpenAI says the result may be inflated by contamination from historical vulnerabilities; GPT-6 Sol 81.7%, GPT-6 Astra 100%. ExploitBench Internal Port, arbitrary code execution success rate: 21.5% (vendor-card; OpenAI GPT-6.1 Sol addendum to the GPT-6 Astra system card, 29 September 2026; OpenAI internal port of vulnerabilities disclosed June to August 2026; safeguards: safeguard state during the run unstated; OpenAI ran its evaluations in its research environment or through the API; run 2026-09-29; CONFIRMED). GPT-6 Astra 31.5%, GPT-6 Sol 5.5%, GPT-5.6 Sol 3.5%. SEC-Bench Pro, pass@1: 78.8% (vendor-card; OpenAI GPT-6.1 Sol addendum to the GPT-6 Astra system card, 29 September 2026; SEC-Bench Pro, vulnerability discovery in JavaScript engines including V8 and SpiderMonkey; safeguards: safeguard state during the run unstated; OpenAI ran its evaluations in its research environment or through the API; run 2026-09-29; CONFIRMED). GPT-6 Astra 85.4%, GPT-6 Sol 66.3%, GPT-5.6 Sol 79.1%. ExploitGym, intended-vulnerability success rate per attempt: 35.1% (vendor-card; OpenAI GPT-6.1 Sol addendum to the GPT-6 Astra system card, 29 September 2026; ExploitGym; safeguards: safeguard state during the run unstated; OpenAI ran its evaluations in its research environment or through the API; run 2026-09-29; CONFIRMED). GPT-6 Astra 42.4%, GPT-6 Sol 22.1%, GPT-5.6 Sol 30.3%. OpenAI cybersecurity safety evaluation, production chat score: 0.987 (vendor-card; OpenAI GPT-6.1 Sol addendum to the GPT-6 Astra system card, 29 September 2026; OpenAI production-chat cyber safety set (Table 15); safeguards: safeguard behaviour measured; higher is better; run 2026-09-29; CONFIRMED). GPT-6 Astra 0.970, GPT-6 Sol 0.957; synthetic agentic environments 0.997, semi-synthetic 0.980; a different metric from the 91.5% refusal figure on the GPT-6 Astra row. Refusal gate (CONFIRMED, observed 2026-10-02): The same safeguards stack as GPT-6 Astra; OpenAI is phasing advanced cyber access to GPT-6.1 Sol through Daybreak. OpenAI says its evaluations ran in its research environment or through the API, which may differ from production ChatGPT. GPT-6.1 Sol shipped to ChatGPT Work and Codex for paid plans and to the API as gpt-6.1-sol on 29 September 2026.