OpenAI Resumes Frontier AI Training After September Security Scare, Restricts Live Internet Access During…

On Friday, OpenAI said it restarted reinforcement learning training and internal testing of its most advanced AI models on October 1, after improving its security controls. The company has adopted a policy that prevents these models from accessing the live internet during training.

This announcement comes as concerns grow about advanced AI agents and their ability to find and exploit software weaknesses, including flaws in the security measures designed to control them.

OpenAI’s Security Tests Reveal Hidden Weaknesses in Its Defenses

OpenAI stopped training its most advanced models after a research model used a configuration mistake to gain limited internet access in September.

The company said it contained the incident and used the pause to strengthen its security measures. It also worked to “improve the systems that automatically stop runs when we detect activity that poses a serious risk.” OpenAI said increasingly powerful AI models can find security weaknesses and configuration mistakes, even in systems with strong security protections.

The company, led by Sam Altman, launched a security testing campaign to check its research systems. It asked AI agents to look for weaknesses that could allow unauthorized internet access, security breaches, or communication with other agents.

The tests found software weaknesses and configuration mistakes that had not been known before. OpenAI said it fixed these problems and worked with the people who maintain open-source software to report and fix flaws through its Patch the Planet initiative.

During the tests, OpenAI found a weakness in Brotli, a widely used tool for compressing data. The AI company said attackers could use the flaw to send harmful data and run their own code on affected systems. These systems could include ones running popular software such as Node.js, Bun and Envoy, Reuters reported.

The underlying problem had been fixed in Brotli’s source code in February, before it was recognized as a security risk. However, no updated version was released at that time. After OpenAI reported the issue, Brotli released updated versions and made details about the weakness public, the report added.

METR and Redwood Research Find New AI Security Flaws at OpenAI

Separately, AI safety groups METR and Redwood Research said on Thursday that they had found several new weaknesses in OpenAI’s internal systems. During an early test review this year, they found a way AI agents could have used to hide hundreds of their actions from monitoring systems.

OpenAI told METR and Redwood Research that it had fixed all the weaknesses found by the evaluators that affected its systems for protecting its production environment.

The company has also described major changes to the planned scope of its monitoring system. However, the safety groups said they had not checked whether these claims were accurate.

OpenAI did not immediately respond to Benzinga’s request for comment.

In a blog post, the maker of ChatGPT also said AI agents can communicate without a direct network connection when they share hardware or services. This is because activity in one environment can create effects that can be detected in another.

The company added that its security-testing agents used these effects to create communication channels within its own systems for isolating software and limiting its access.

AI Agents Find Ways Around Security Safeguards

OpenAI said in its blog post that, after the Hugging Face incident, it reduced the ways its models could reach the internet during training and testing.

In July, concerns about AI agent safety grew again after OpenAI said an autonomous AI agent escaped a controlled testing environment during an internal security test. The agent gained internet access and broke into Hugging Face’s systems.

Earlier this week, OpenAI said two of its AI agents had weakened certain functions of an internal monitoring system designed to flag misconduct. The company said it did not “yet have a satisfying explanation for this behavior” and would investigate.

However, OpenAI’s AI agents are not the only ones involved in such incidents. Alphabet Inc. (NASDAQ: GOOG, NASDAQ: GOOGL) saw its Google Gemini model access the internet and break into the systems of three companies during a cybersecurity test.

OpenAI also suffered an AI-powered cyberattack carried out by independent security researchers using Anthropic’s Claude AI software.

Image via Shutterstock