OpenAI classifies GPT-6.1 Sol as Critical in cybersecurity under its Preparedness Framework. On one internal test of recently disclosed software vulnerabilities, the model recorded a 21.5% arbitrary-code-execution rate, meaning it succeeded in getting vulnerable software to run code of its choosing. GPT-6 Sol scored 5.5% and GPT-6 Astra 31.5% on the same test.
“Critical” is OpenAI’s capability classification, not an incident label. Its framework reserves the tier for models able to develop functional zero-day exploits across hardened real-world systems without human intervention, or to devise end-to-end attacks against hardened targets from a high-level goal. OpenAI says the classification rests on automated cyber evaluations; this score is one result among several reported in the addendum.
The Internal Port evaluation tests whether models can develop working exploits for recently disclosed vulnerabilities, reducing the chance that results reflect exposure to older examples. OpenAI describes GPT-6.1 Sol’s result as a substantial increase over earlier Sol models but below Astra, and says reliable exploitation remained challenging for Sol on this test.
OpenAI says it applies the same safeguards stack to GPT-6.1 Sol as to Astra. The figures are company-run benchmark results, not an independent replication or an estimate of real-world attack success; they describe performance on the reported tests, not how often the model would succeed against targets in the wild.
Sources
- OpenAI: Addendum to GPT-6 Astra System Card — GPT-6.1 Sol, published Sept. 29, 2026.
- OpenAI: GPT-6 Astra System Card, framework definitions and comparison context.
- Illustrative image: Negative Space, Laptop Programmcode.jpg, via Wikimedia Commons, CC0 1.0.
