OpenAI launches always-on AI agents a day after apologizing for a hack by its bots

Questions of trust and safety hovered over the event. The evening before the new products were launched, OpenAI apologized for its agents’ hacking of an Australian government website. “We are sorry and working to do better in the future,” it said in a blog post.

The incident, which occurred in June, involved an internal model under development, which accessed nonpublic information related to Australia’s public health insurance program. On Tuesday, Altman and OpenAI leadership stressed that the episode — and other major incidents of OpenAI’s systems’ going rogue or acting in unintended ways — all involved models that were not meant for public release and were distinct from the commercial offerings available to users.

“The instances we disclosed were running internal-only models that were not on track to ship,” said Alex Embericos, a product lead at OpenAI.

Despite the distinction, Altman said there is a trade-off between pushing the frontier of AI development and building cutting-edge AI systems safely.

“To push models to the new level of capability, it is going to take us some time to figure out how to make sure that alignment, monitoring, safety and security stay well ahead of capabilities,” he told reporters after his keynote speech. “But you can see that we were able to build a huge amount of useful new stuff without a major new model, and there will be new models.”