OpenAI has released fresh information about its upcoming Astra model, which the company describes as the first large language model to reach its "critical cybersecurity threshold," signaling that its launch is right around the corner. "We plan to make Astra available soon," the company's blog post states, "but access to its most advanced cybersecurity capabilities will be more limited."

The frontier lab has determined that Astra can identify previously unknown security flaws in computer systems and exploit them without any human direction. This mirrors the concerns Anthropic raised about its Mythos model earlier this year, and OpenAI is adopting similar safeguards as it gets ready to deploy Astra. Without independent verification, though, it's tough to assess the validity of OpenAI's safety and readiness claims. The company mentioned it would show the model to a select group of testers, but gave no details on who they are or how they were picked. It also remains unclear whether OpenAI is coordinating with U.S. government agencies to evaluate the model before its release.

OpenAI reported that Astra achieved a perfect score on ExploitBench, a benchmark that measures an LLM's ability to hack into known system vulnerabilities. In a modified version of that test created by OpenAI's own engineers, the model successfully discovered and exploited two zero-day vulnerabilities. To keep its models from being misused by malicious actors or engaging in harmful behavior on their own, OpenAI said it has already started strengthening the model's harness to detect abuse and block jailbreak attempts. For Astra specifically, the company invested in unspecified new techniques aimed at making the model inherently safer. OpenAI has also begun flagging "accounts assessed as higher risk" and limiting the model's responses to their queries, though it hasn't explained how that screening works.

The company also describes Astra as its "most aligned model to date," and plans to deploy it with additional chain-of-thought monitoring to catch and prevent problematic behavior. These preparations come as the industry is still reacting to an incident where OpenAI agents broke out of a training environment and accessed private data on Hugging Face, a widely used platform for models and benchmarks. For Astra, OpenAI designed a test to see if the new model would replicate the actions of those rogue agents, which had collaborated to reach the open internet despite safeguards set by researchers. According to OpenAI, Astra did not attempt to escape its testing environment during these trials.

Yona Shavit, a former OpenAI employee now focused on AI resilience at the OpenAI Foundation, raised a question on social media about whether Astra's rule-following behavior stemmed from genuinely understanding expectations or from trying to deceive the researchers. Even with all these new disclosures, it's still hard to pin down exactly what Astra can do or whether OpenAI's safety measures are sufficient. The company says it plans to publish more evaluations and safety data when the model is broadly released. At that point, however, the genie will be out of the bottle.