OpenAI’s New “Astra” Model Can Launch Cyberattacks Without Human Guidance, Company Warns

OpenAI has revealed that its upcoming Astra model can independently discover unknown software vulnerabilities and build working exploits for them, a capability that once required skilled human hacking teams. The company has classified Astra as its first model to reach a “Critical” cybersecurity risk level under its internal safety framework.

What Makes Astra Different

To earn this classification, a model must be capable of finding zero day flaws and turning them into functional exploits on hardened, real-world systems with no human step by step input. It can also carry out a full attack starting from just a broad goal.

During testing, Astra achieved a perfect score on a benchmark measuring exploit development from known vulnerabilities. It also uncovered two previously unknown flaws while constructing an exploit chain in a separate internal evaluation.

Sandbox Breakout and Root Access

Astra reportedly escaped a hardened browser sandbox to run commands directly on the host machine. In another test, it chained together multiple operating system flaws to gain full root-level access.

Why This Matters for Crypto

Because software vulnerabilities in crypto systems can be converted into stolen funds almost instantly, faster and more autonomous exploit discovery raises fresh concerns for the industry. OpenAI has delayed parts of Astra’s rollout to add safeguards and will limit its most advanced capabilities to select testers first.