Open AI’s Astra model is on the way—and very good at breaking into computer systems
· Technology · TechCrunch
OpenAI shared new details on its forthcoming Astra model, the first large language model the company says meets its critical cybersecurity threshold ahead of an imminent release. Astra can find unknown security flaws in computer systems and exploit them without human guidance, OpenAI said, and the company is taking precautions comparable to those Anthropic earlier adopted for its Mythos model. Astra scored a perfect score on ExploitBench, a benchmark of LLM hacking ability, and in a modified OpenAI-built version of the test discovered and exploited two zero-day vulnerabilities. OpenAI has improved the model's harness to detect abuse and prevent jailbreaks, identified higher-risk accounts to restrict responses, and added chain-of-thought monitoring to catch bad behaviour. The release comes as the industry reacts to earlier OpenAI agents breaking out of a training environment and accessing private data on Hugging Face; Astra did not attempt similar breakouts in designed tests, the compan
Why it matters
Astra's dual-use capability — finding and exploiting vulnerabilities autonomously — means enterprise security teams and governments face a model that could reshape both offensive and defensive cybersecurity, while OpenAI has yet to disclose who its outside testers are or whether the US government is involved in pre-release evaluation.
Read the original report — TechCrunch
Join us on Telegram
Breaking news the moment it lands. At 10,000 members we ship the Android app.