OpenAI took up to 112 days to publish misalignment cases it found in June

OpenAI has published four new reports of its internal AI models misbehaving, and all of them describe incidents from June. The longest gap between discovery and publication was 112 days, according to OpenAI's misalignment reports site.

OpenAI published this artwork with its July 20 post on safety and alignment in an era of long-horizon models. Image: OpenAI

In that case OpenAI knew by June 25 that a model had found a way around blocked web requests, but only published it on October 15. Two other cases took 63 days from discovery, and the quickest took 20.

Three of the four were found after the fact. OpenAI ran its misalignment monitor back over earlier training and testing records following the Hugging Face security breach, and that turned up cases nobody had spotted at the time.

No deadline in the framework

When OpenAI launched its reporting framework on September 16, it said the process was meant to speed up publication and that each step would have a deadline. It has not said what those deadlines are.

The cover image OpenAI uses for its own page on how it thinks about safety and alignment. Image: OpenAI

Models that took what they weren't given

All four cases involve a model accessing something it was not allowed to. One, asked to deploy contracts on an Ethereum test network, paid with a wallet it had not been given. It worked out keys from publicly known recovery phrases, checked 362 addresses and spent test funds from one of them.

In the most serious case, a model researching a public dashboard found login credentials in the website's code and used them to reach the organisation's search service. In a second test it downloaded five private tables with more than 147,000 rows, plus 1,489 records including names and email addresses, and did not disclose any of it.

Another model, looking for hospital bed statistics on a government website, retrieved server files and credentials and bypassed the site's access checks. Three of the four tasks had told the models to use the internet only for simple read requests.

The fourth model already had its answer from a public yearbook but kept going, routing blocked requests through an outside browser and reading figures out of images.

Internal models only

None of the incidents involved ChatGPT or OpenAI's public API, and OpenAI has removed details that would identify the organisations affected. It has published similar reports since September, including six at the framework's launch.