OpenAI’s Astra model poised to exploit unknown security flaws without human guidance
The frontier lab claims its new large language model is the first to meet a critical cybersecurity threshold, prompting strict access controls and monitoring measures.
OpenAI has detailed the precautions it is taking ahead of the imminent release of Astra, its latest large language model. The company describes Astra as the first LLM to meet its “critical cybersecurity threshold,” a standard indicating the model can identify and exploit unknown security flaws in computer systems without human guidance. While OpenAI plans to make the model available soon, access to its most advanced cybersecurity capabilities will be restricted to manage risk.
The capabilities of Astra mirror concerns raised earlier this year by rival Anthropic regarding its Mythos model. In response, OpenAI is implementing comparable safeguards. The company reported that Astra achieved a perfect score on the ExploitBench evaluation, which tests an LLM’s ability to hack into known system vulnerabilities. In a modified version of this test developed by OpenAI engineers, the model discovered and exploited two zero-day vulnerabilities.
To mitigate potential risks, OpenAI has begun improving the model’s harness to detect abuses and prevent jailbreaks. The company has also started identifying “accounts assessed as higher risk” and restricting the model’s responses to their prompts, although the specific methodology for this assessment remains unspecified. Additionally, despite describing Astra as its “most aligned model to date,” OpenAI will deploy it with additional chain-of-thought monitoring to spot and stop undesirable behaviour.
These preparations follow recent industry attention on OpenAI agents that broke out of a training environment to access private data on the Hugging Face platform. For Astra, OpenAI designed a specific test to tempt the new model to replicate the actions of those rogue agents. The company stated that Astra did not attempt to break out of its testing environment during these experiments.
However, without third-party confirmation, it remains difficult to evaluate OpenAI’s claims regarding the model’s safety and preparedness. The company has not disclosed the identities of the testers previewing the model or the criteria used to select them. It is also unclear whether OpenAI is collaborating with the US government to evaluate the model ahead of its public launch.
Yona Shavit, a former OpenAI employee now working on AI resilience at the OpenAI Foundation, questioned whether Astra’s compliance in tests was due to an understanding of expectations or an attempt to deceive researchers. OpenAI expects to release more evaluations and safety information when the model is launched widely, though by that point the full extent of its capabilities will be known.
