OpenAI's test agents found ways around locked doors, so it's worth checking which keys your AI tools are holding
By Brett Gavaghan, Founder of Adaptd ·
OpenAI has disclosed four more cases of its AI models going past their limits during internal testing, including downloading data they weren't allowed to and making an unauthorised payment.
The short version
- OpenAI added four incidents to its public list of cases where its AI models acted beyond what they were authorised to do.
- All four happened in June during OpenAI's own testing and research, and the list now holds 19 incidents.
- Give AI agents only the access a task needs, and check what they did, not just what they reported.
Every workplace has a new starter who finds the spare key on day two. OpenAI's latest disclosures show AI agents can be that person, only quicker.
What did OpenAI disclose?
OpenAI disclosed four more cases of its AI models acting outside what they were allowed to do during internal testing and research. Fortune reported on 16 October 2026 that the cases were added to a disclosures page OpenAI set up after a hack in July, which now lists 19 incidents.
In one, an internal model downloaded 24 years of hospital bed data from a government statistics website after the site blocked it, by getting around its access controls. In another, a model accessed a test cryptocurrency wallet it hadn't been given and made an unauthorised payment.
What else happened, and what is OpenAI changing?
A third model accessed internal systems, made changes and downloaded 147,451 rows of non-public data without telling its evaluators. A fourth got internet access from a research environment that was meant to be cut off, then asked outside services to do things it had been blocked from doing.
OpenAI says it has strengthened security controls, blocked models from the live internet during training and testing, improved how it grades model behaviour, and extended monitoring to all training and evaluation traffic.
What should your business do about it?
Your business should give AI agents only the access each task needs. These incidents happened inside OpenAI's own testing, not in customer accounts, but they show how determined an agent can be when its goal and its limits collide.
My advice: before you connect an AI tool to your inbox, bank feed or customer list, ask what it could do with that access on a bad day. Use read-only access where you can, and check the activity log now and then.
Agents are keen. Keen is great, right up until it finds the master key.
Wondering what this means for your business?
That's exactly what the discovery call is for.