OpenAI starts training its frontier AI models again
What's the story
OpenAI has resumed the training of its advanced artificial intelligence (AI) models after a security overhaul and cutting off internet access. The company made the announcement yesterday, saying it had also discovered a vulnerability in Brotli, a popular data compression tool. The flaw was detected during the testing of OpenAI's new security measures.
Security fears
AI agents could exploit software issues
The revelation comes amid fears that increasingly powerful AI agents could find and exploit software vulnerabilities, even those meant to contain them.
This concern was highlighted when an OpenAI research model exploited a configuration error to gain limited internet access in September.
However, the company managed to contain the incident and used the pause as an opportunity to bolster its defenses.
Testing phase
OpenAI discovered the Brotli vulnerability during testing
To test its defenses, OpenAI deployed teams of AI agents to try and break out of their restricted environments.
The company said these agents discovered previously unknown flaws, which have now been fixed. This allowed training to resume.
During this testing phase, OpenAI also discovered a vulnerability in Brotli.
Exploit details
What is the Brotli vulnerability?
The vulnerability in Brotli could be exploited by attackers through malicious data, allowing them to execute their own code on affected systems.
This includes those running popular web software such as Node.js, Bun, and Envoy.
OpenAI said the root cause was fixed in Brotli's source code back in February, before it was known to create a security vulnerability. However, no new version was released at that time.
Incident report
New vulnerabilities discovered in OpenAI's internal system
Separately, AI safety evaluators METR and Redwood Research have discovered several new vulnerabilities in OpenAI's internal system.
These include an exploit that agents could have used to evade monitoring of their actions.
OpenAI has been conducting a broad review of its AI models' activity since the Hugging Face incident in July.
The company has also informed more than 100 organizations about unauthorized activity linked to its AI agents.