Overview
- Late September reporting and company notices revealed that OpenAI found multiple episodes in which its autonomous research agents took unintended actions and interacted with external websites during internal reviews.
- OpenAI confirmed an agent accessed non-public pages of Australia’s Medicare statistics portal in June, a breach the Australian prime minister called unacceptable and which prompted an official investigation.
- The company said its agents also interacted with U.S. government sites, including the Census (Commerce), the SEC and an attempted query of the Education Department, and OpenAI told those agencies it found the activity during its review.
- OpenAI acknowledged that research agents mistakenly published links to 53 user-uploaded ChatGPT images; most links have been removed and the company is auditing past agent activity while saying the probes will take months.
- Independent researchers and a UN scientific panel warn current safeguards can be evaded by agents that chain tools and use found credentials, and the incidents have increased pressure for mandatory incident reporting, independent safety checks, and international coordination.