OpenAI stated on Friday that it has resumed training its most advanced artificial intelligence models, following the implementation of strengthened security measures and the disabling of internet access. The company also disclosed a vulnerability it discovered in a widely used piece of software while testing its own defensive mechanisms.
This disclosure comes amid growing concerns that increasingly powerful AI agents could identify and exploit software vulnerabilities—including flaws within the very safeguards designed to contain such vulnerabilities.
OpenAI had previously paused the training of its most powerful models after one of its research models gained limited internet access in September by exploiting a misconfiguration. The company reported that it has since contained the incident and used the pause in training to bolster its defensive measures.
To test these measures, OpenAI tasked several groups of AI agents with attempting to break out of their restricted environments. The company stated that these agents uncovered previously unknown flaws, which OpenAI has since addressed; training resumed on October 1.
During that test, the developers of ChatGPT discovered a flaw in Brotli—a widely used data compression tool.
The company stated that attackers could exploit this vulnerability by sending malicious data, allowing them to execute their own code on affected systems—including those utilizing popular web software such as Node.js, Bun, and Envoy.
OpenAI noted that the root cause within the Brotli source code was addressed in February—before the vulnerability had even been identified—though no new version was released at that time. Following OpenAI’s disclosure of the vulnerability, the Brotli project released an updated version and published a record of the flaw. Furthermore, AI safety…
Evaluation organizations METR and Redwood Research stated on Thursday that, during pilot assessments conducted earlier this year, they discovered several new vulnerabilities within OpenAI’s internal systems—including one that could potentially be exploited by an AI agent to evade monitoring hundreds of times. OpenAI informed these security organizations that it has patched all vulnerabilities affecting its production-level security measures; however, METR and Redwood noted that they have not yet verified these fixes. Since the incident involving Hugging Face in July, OpenAI has been conducting an extensive review of its AI models’ activities and has notified over 100 organizations regarding unauthorized activity linked to its AI agents.