OpenAI unveils faster Astra model while warning it can evade monitoring

OpenAI has unveiled GPT-6 Astra, a faster artificial intelligence model that the company calls its best yet. But alongside its capabilities, OpenAI disclosed a problem that could make the system harder to supervise: Astra sometimes tries to conceal or disguise how it reaches an answer.

The model is designed to handle tasks with little human intervention. OpenAI said it can prepare taxes, develop games, produce architectural renderings, format legal memorandums and search for apartments. In one example, the company said Astra reduced cat-sitter research from 30 minutes for a human to five minutes. It said a job search took two minutes and 51 seconds with Astra, compared with five hours without it.

The disclosure comes as OpenAI faces scrutiny over the behavior of its agents. In July, agents escaped a secure test and hacked into the open-source platform Hugging Face while attempting to cover their tracks. Similar concerns have emerged at rival Anthropic as developers race to deploy increasingly autonomous systems.

OpenAI said Astra is not yet consistently able to obscure its methods on more complicated problems, although its ability to conceal its tracks is improving. The company’s chief scientist said at a Thursday briefing that monitoring agents was becoming more difficult, as was alignment, the effort to ensure that AI systems reflect human values.

Monitoring is central to OpenAI’s assurances to regulators, lawmakers and the public after the security incident. The company also told two Democratic members of the U.S. House this week that it is developing “automated shutdown capabilities” for its models.

“

automated shutdown capabilities