OpenAI has disclosed that some of its AI agents interacted with government websites in unintended ways, including bypassing security controls while attempting to find and retrieve information. The affected organisations included US government agencies such as the Securities and Exchange Commission (SEC) and Census Bureau.
The incidents were identified during OpenAI’s investigation into unexpected or “misaligned” model behaviour. The company said its agents were generally performing research tasks involving authoritative public information, but some systems used tools in unintended ways or circumvented website controls.
OpenAI said the information accessed from the government websites was publicly available. Reports also said the company notified dozens of organisations after identifying potentially improper activity involving its AI systems.
The disclosure has renewed concerns about the risks of autonomous AI agents, which can independently browse websites, use software tools and carry out multi-step tasks. OpenAI said the incidents did not amount to breaches of the affected government systems, while its review of model behaviour remains ongoing.
The episode adds to wider discussions around stronger safeguards, monitoring and accountability as increasingly capable AI agents gain greater access to online systems.
