OpenAI has temporarily halted training of its latest AI models following reports that its agents took actions beyond their instructions while accessing federal government websites, according to a report fropm the Associated Press (AP). This is only the most recent time OpenAI and other frontier AI labs have seen agents it is they are testing break free of their constraints and act on the open internet.
The company is reviewing several incidents from this summer involving agents that were gathering information from government sites. In one case involving the Securities and Exchange Commission, the report said, agents accessed publicly available information but then posted it elsewhere online, an action they had not been instructed to take. The SEC says no nonpublic information was accessed.
In another incident, agents discovered API developer keys while accessing a Department of Education website, although OpenAI says they ultimately gathered only publicly available information. The department says it found no evidence its website or databases were affected.
Separately, AI evaluator Transluce reported that agents believed to be from OpenAI unsuccessfully attempted to hack into a Department of Education website. OpenAI has not confirmed that report, according to the AP.
“We will resume training only when we are confident that we have additional safeguards,” the company said.
The move comes as AI developers face increasing scrutiny over the ability of autonomous agents to take actions across websites and enterprise systems. While the agentic AI that powers most enterprise automation systems is not the same as those agents being tested by the frontier labs, security threats that undermine agentic AI of any kind could affect attitudes toward the technology.