OpenAI flaws could let AI agents evade action monitoring
What's the story
Two non-profit groups that check AI safety, METR and Redwood Research, have found security weaknesses in OpenAI's internal system for watching what AI does. The groups said on Thursday that they found these problems during a test review held from July 21 to August 11. As part of OpenAI's ongoing work with outside AI safety groups, the researchers were allowed to look closely at the company's technology and internal systems. They described this as getting "a deep level of access" to OpenAI's infrastructure.
Access details
Outside researchers were allowed to examine OpenAI's internal systems
During the test review, researchers from outside OpenAI were allowed to examine the company's internal systems.
Each of the two organizations sent one researcher. Both researchers were allowed to see OpenAI's internal computer code, technical documents, and some of its technical employees.
According to a statement from METR and Redwood Research, the researchers were supervised while working at OpenAI's offices. They used devices provided by OpenAI, and those devices "never left the premises."
Fixes implemented
Researchers found new kinds of security weaknesses
The researchers found several security weaknesses they described as "novel vulnerabilities," meaning problems that had not been identified before. One of these weaknesses could have allowed AI agents to hide "hundreds" of their actions from the systems meant to watch them.
However, OpenAI says it has fixed all the security weaknesses that affected the safety protections used in its live systems.
The researchers gave OpenAI a private report that was 61 pages long. But they have not yet checked for themselves whether OpenAI's fixes work.
System assessment
The test checked OpenAI's system for spotting unwanted AI behavior
During the three-week test, the researchers pushed OpenAI's internal monitoring system to its limits. This system is designed to spot when AI behaves in ways that do not match what it is supposed to do.
They tested the system as it worked on July 21 and before that date. They also pointed out that OpenAI was making the system bigger and adding to it while the tests were happening.
This review is part of a larger effort by AI companies, experts, and even the Trump administration to encourage independent safety checks by outside experts. The goal is to help make AI technology safer without slowing down its development.
Information security
AI companies worry that secret information could reach competitors
Although independent safety checks can be helpful, AI companies are worried about allowing outside researchers to work inside their offices and use their internal systems.
They fear that these researchers could accidentally or deliberately share private company information with competing businesses.
Reuters reported on these worries earlier this month. The issue shows how difficult it can be for AI companies to be open about their technology while also keeping important company information safe.