← Back to feed

OpenAI released the safety card addendum for GPT-6.1 Sol on September 29, treating it as a Critical capability in Cybersecurity under the Preparedness Framework. Critical is defined as a model that can “identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention” or “devise and execute end-to-end novel strategies for cyberattacks against hardened targets given only a high level desired goal.” GPT-6 Astra received the same designation. GPT-6 Sol and GPT-6 Luna, released September 22, were rated High rather than Critical. GPT-6.1 Sol is the first model in the Sol line to reach the Critical tier.

ExploitBench numbers

ExploitBench evaluates whether a model can turn known vulnerabilities into increasingly powerful exploit primitives — up to arbitrary code execution — given vulnerability descriptions, source code, and patches, without a reference exploit. At maximum reasoning effort, GPT-6.1 Sol scores 99.7%. GPT-6 Sol scored 81.7%. GPT-6 Astra scored 100%.

An internal evaluation called ExploitBench Internal Port — designed to test exploitation of recently disclosed vulnerabilities with lower contamination risk — shows GPT-6.1 Sol at 21.5%, compared to roughly 5% for GPT-6 Sol. The jump from 5% to 21.5% on recently disclosed vulnerabilities is more operationally significant than the ExploitBench headline: it tests real exploitation of current software issues, not known CVE training data.

Bio and Chemical: High, not Critical

GPT-6.1 Sol earns a High capability rating in the Biological and Chemical domain. High means the model can provide meaningful uplift to someone attempting to cause mass-casualty events using these domains. OpenAI reports that Sol’s results did not cross the indicative Critical thresholds in this category.

AI Self-Improvement: below threshold

In AI Self-Improvement evaluations, GPT-6.1 Sol does not cross either the High or Critical threshold.

Production robustness

On OpenAI’s production benchmarks, GPT-6.1 Sol scores higher than GPT-6 Sol in five of eight categories. All GPT-6 series models outperform earlier models on both static and multi-turn robustness evaluations. GPT-6 Luna shows some higher individual scores, but the safety card notes this may partly reflect a tendency to refuse legitimate requests rather than genuine robustness gains.

HealthBench

At length-adjusted scoring, GPT-6.1 Sol reaches 64.2 on HealthBench Professional (+3.4 relative to GPT-6 Sol), 58.5 on HealthBench (+5.3), 36.2 on HealthBench Hard (+6.1), and 96-point-something on HealthBench Consensus (down 0.2). Those figures put Sol within 0.5 points of GPT-6 Astra across all four HealthBench evaluations.

The pattern

Two models in the GPT-6 family — Astra and now Sol — have both received Critical cybersecurity ratings. The same week that OpenAI published this addendum, Anthropic published research finding that GLM-5.3, an open-weight Chinese model, achieved near-frontier offensive cyber capability with safeguard bypass rates of 64-100%. The convergence of Critical ratings at the frontier and near-Critical capability in the open-weight tier represents a compressed window for both defensive preparedness and policy response.