OpenAI resumes frontier AI training after security overhaul, discloses flaw in widely used software

OpenAI has resumed training its most advanced AI models following a security overhaul and a discovery of a flaw in a widely used software, Brotli. The company tightened security measures and barred internet access to its systems after a research model gained limited internet access in September. To test the new defenses, OpenAI employed AI agents to attempt breaking out of their restricted environments, leading to the discovery of unknown flaws. These flaws were subsequently fixed, and training resumed on October 1.

The ChatGPT maker found a vulnerability in Brotli, a data compression tool used by various web software such as Node.js, Bun, and Envoy. Attackers could exploit this flaw by sending malicious data that allows them to execute their own code on affected systems. OpenAI disclosed this issue and stated that the root cause was fixed in Brotli's source code in February, but no new version was released at that time.

Additionally, AI safety evaluators METR and Redwood Research recently discovered several new vulnerabilities in OpenAI's internal system, including one that could prevent agents from having their actions monitored. OpenAI has since patched all vulnerabilities affecting its production safeguards. However, METR and Redwood have not yet verified the effectiveness of these fixes.

OpenAI is currently conducting a comprehensive review of its AI models' activity in the wake of the Hugging Face incident in July and has informed more than 100 organizations about incidents of unauthorized activity involving its AI agents.