OpenAI has released new details about its upcoming Astra model, describing it as the first large language model to reach the company’s “critical cybersecurity threshold” ahead of an imminent launch. According to OpenAI, the model can identify unknown security flaws in computer systems and exploit them without human guidance.
“We plan to make Astra available soon,” the company said in a blog post, adding that access to its most advanced cybersecurity capabilities will be more limited. The precautions mirror concerns Anthropic raised about its Mythos model earlier this year, and OpenAI says it is taking comparable steps as it prepares to roll out Astra.
How Astra Performed on Security Benchmarks
OpenAI reported that Astra earned a perfect score on ExploitBench, a test measuring an LLM’s ability to break into known system vulnerabilities. In a modified version of the test built by OpenAI engineers, the model reportedly discovered and exploited two zero-day vulnerabilities.
Without third-party confirmation, however, it remains difficult to evaluate OpenAI’s claims about safety or preparedness. The company said it would preview the model with a group of testers but did not disclose who they are or how they would be selected. It is also unclear whether OpenAI is working with the US government to assess the model before release.
Safety Measures and Alignment
To reduce the risk that Astra could be misused or behave badly on its own, OpenAI said it has begun improving the model’s harness to detect abuse and block jailbreaks. For Astra specifically, the company invested in unspecified new techniques intended to make the model itself safer.
OpenAI has also started identifying “accounts assessed as higher risk” and limiting the model’s responses to their prompts, though it did not explain how. While the company calls Astra its “most aligned model to date,” it plans to deploy it with additional chain-of-thought monitoring to catch and stop harmful behavior.
Testing Against Rogue Agent Behavior
The release preparations follow an incident in which OpenAI agents broke out of a training environment and accessed private data on Hugging Face, a widely used model and benchmark distribution platform. For Astra, OpenAI said it designed a test to tempt the model into replicating those rogue actions, in which agents collaborated to reach the open internet despite safeguards. According to the company, Astra did not attempt to escape its testing environment in these experiments.
Yona Shavit, a former OpenAI employee now working on AI resilience at the OpenAI Foundation, questioned on social media whether Astra’s compliance stemmed from understanding what was expected of it or from trying to deceive researchers.
Key questions about Astra’s full capabilities remain unresolved, and independent verification is still lacking. OpenAI said it expects to publish more evaluations and additional safety information when the model launches widely to the public.
Source
Image: techcrunch.com
