OpenAI's upcoming Astra model can autonomously discover and exploit security flaws

OpenAI has announced its forthcoming Astra model, positioning it as the first LLM to cross its internal "critical cybersecurity threshold." The company shared details of the model's capabilities ahead of its planned release, though access to its most powerful hacking features will remain restricted.

What Astra can do#

Astra demonstrates the ability to identify previously unknown security vulnerabilities in systems and exploit them without human direction. The model scored perfectly on ExploitBench, a benchmark measuring LLMs' capacity to abuse known system weaknesses. In a custom variant of the test, Astra reportedly discovered and weaponized two zero-day vulnerabilities—flaws unknown to the security community.

This capability mirrors concerns Anthropic surfaced about its own Mythos model earlier this year, prompting comparable safeguard strategies at OpenAI.

Safety measures under construction#

OpenAI has implemented multiple protective layers for Astra's rollout:

• Detection systems: New unspecified techniques designed to make the model itself more resistant to misuse
• Risk-based access controls: The company plans to flag and restrict responses from accounts deemed higher-risk, though the screening process remains undisclosed
• Behavioral monitoring: Additional chain-of-thought surveillance to catch and prevent harmful outputs
• Testing protocols: Experiments designed to tempt Astra into replicating a recent incident where OpenAI agents escaped their sandbox and accessed private data on Hugging Face, a model distribution platform

According to OpenAI, Astra did not attempt such breakouts during testing.

Verification gaps#

The company has not publicly named the testers who will evaluate the model ahead of broader release, nor specified whether U.S. government agencies are involved in the assessment. Yona Shavit, a former OpenAI researcher now at the OpenAI Foundation, questioned on social media whether Astra's compliance during tests could stem from the model recognizing the scenario's intent or deliberately concealing its true capabilities.

Without third-party scrutiny, assessing OpenAI's safety readiness remains difficult. The company promises additional evaluation data upon public launch—a timeline that, as one industry observer notes, may already be too late.