Astra reportedly achieved a perfect score on ExploitBench, a benchmark designed to evaluate AI’s ability to identify and exploit real-world vulnerabilities. This places the model at the forefront of AI-driven security research, but it also raises concerns about dual-use potential. In response, OpenAI has moved to block prompt requests that attempt to generate proof-of-concept exploit code, effectively restricting the model’s offensive security output.

Beyond exploit generation, the model is said to be state-of-the-art in computer use, browsing, and software engineering tasks, making it a versatile tool for both defensive and offensive security workflows. The "Critical" rating under the Preparedness Framework signals that OpenAI will continue to monitor deployment closely and may impose usage restrictions in high-risk scenarios.

• GPT‑6 Astra achieved 100% on ExploitBench.
• OpenAI blocked PoC exploit generation requests.
• Model crossed "Critical" cybersecurity threshold pre-launch.
• Excels in computer use, browsing, and software engineering.

The decision to restrict exploit code requests is notable, as it acknowledges the model’s capability to produce actionable attack tools while attempting to maintain a safety boundary. Whether these restrictions are enforceable in practice remains an open question for the security community.

Source: The Hacker News

Is your team already testing or evaluating GPT‑6 Astra for defensive use, and how are you balancing its offensive potential against policy restrictions?