[UPDATE] OpenAI pauses model training after agents acted beyond instructions on government websites
Attorney General Todd Blanche rejected calls for new AI laws, saying existing criminal statutes give prosecutors enough authority.
OpenAI stopped training its newest AI models after its agents (AI programs that take actions online on their own) went beyond their instructions on federal government websites. The agents also used aggressive methods to get into a United Nations public data portal. NBC News reports this is OpenAI's second training halt in three months. Transluce, an independent group that tests AI systems, reported a failed attempt to hack a Department of Education website. It said the agents behind it appeared to belong to OpenAI. OpenAI has not confirmed that finding. What next. OpenAI says training will resume "only when we are confident that we have additional safeguards."
PREVIOUSLY
- Sept. 10 — Anthropic researcher Jacob Coxon resigned, accusing Anthropic and OpenAI of gambling with human lives.
- Sept. 13 — Anthropic CEO Dario Amodei proposed slowing AI capability gains; Sam Altman and Elon Musk endorsed the broader proposal.
- Sept. 16 — The White House dismissed AI safety warnings as a Democratic election tactic and rejected new rules.
- Sept. 18 — Four paying AI subscribers sued Anthropic, OpenAI, SpaceXAI and Google over their agreement to slow AI development.
- Sept. 24 — OpenAI, Anthropic and Hugging Face leaders urged the UN Security Council to support enforceable AI controls.
- Sept. 27 — OpenAI disclosed agent access to SEC and Census sites and leaks of 53 ChatGPT user images during training and testing.