OpenAI has announced the forthcoming release of its Astra model, which the company claims is the first large language model to meet its “critical cybersecurity threshold.” According to OpenAI, Astra demonstrates the capability to identify and exploit unknown security vulnerabilities in computer systems autonomously, without human guidance. The company reports that Astra achieved a perfect score on ExploitBench, a benchmark for evaluating an LLM’s ability to exploit known system vulnerabilities, and further discovered and exploited two zero-day vulnerabilities in a modified test developed by OpenAI engineers.

In preparation for Astra’s release, OpenAI is implementing several safety measures. The company has begun enhancing the model’s harness to detect and prevent abusive behavior and jailbreak attempts. Additionally, OpenAI is identifying “accounts assessed as higher risk” and restricting the model’s responses to prompts from these accounts, although the specifics of this process are not detailed. OpenAI also plans to deploy Astra with chain-of-thought monitoring to identify and halt any malicious behavior. Despite these precautions, the article notes the difficulty in independently verifying OpenAI’s claims about the model’s safety and the effectiveness of its mitigation strategies, as the company has not disclosed the composition of its tester group or any collaboration with governmental entities for pre-release evaluation. OpenAI intends to provide further evaluations and safety information upon the model’s wider public launch.

This is a summary. Read the original at techcrunch.com.