OpenAI to Launch Astra Model- Breaking into Your Computer

OpenAI has shared new information about its forthcoming Astra model, which the company says is one of the largest language models to meet its critical cybersecurity threshold. According to the Frontier Lab, Astra can find unknown security flaws in computer systems and exploit them without any person’s guidance. 

These capabilities are in line with the concerns raised earlier about the Mythos model at the start of the year. And without third-party approval for safety and preparedness, it will be difficult for OpenAI to evaluate. The company is taking comparable precautions and said it would preview the model with a group of testers but didn’t disclose the members’ names. They also mentioned that if the model isn’t approved, the company will work with the US government and will evaluate the model before release. 

OpenAI also said Astra scored a perfect score on ExploitBench, an evaluation of an LLM’s ability to exploit known system vulnerabilities. In a modified version of the test developed by OpenAI, Astra Model discovered and exploited two zero-day vulnerabilities, the company said.

OpenAI also mentioned improving the model’s harness to detect abuses and prevent jailbreaks to avoid exploitation by bad actors and avoid bad behaviour itself. For Astra, the company invested in some unspecified new techniques designed to make the model safer. They have also started identifying “accounts assessed as higher risk” and restricted the model to respond to their prompts. At the end, the company described the Astra model as its “Most Aligned Model to Date”.

A former OpenAI employee, Yona Shavit, now works on AI resilience at the OpenAI Foundation and wondered on social media whether AI models have rule-following behaviour or are just an artefact of how they are tested.