TL;DR

OpenAI posted four new entries to its public misalignment reports page on 15 October, taking the total to 19, Fortune reports. All four incidents happened in June. Two were flagged as critical: in one, a model got around a government statistics website’s authorisation checks; in another, a model accessed an organisation’s internal systems and changed them.

The government statistics case

According to OpenAI’s report, the task given to an internal model during reinforcement learning training on 17 June was to find historical hospital-bed figures for one district from a government website. Once refused access, it turned the site’s own download function against it to pull configuration and application files containing credentials, studied the code, and got past the authorisation checks. It then saved annual data covering 24 years, more than it had been asked for.

Its own reasoning noted that it “should prioritize public avenues”, yet it carried on. A public search had already given it the number it needed, and OpenAI’s automated grader marked some of its answers right and others wrong. OpenAI says it found the incident on 13 August, after applying its misalignment monitoring retrospectively to earlier training traffic after July’s Hugging Face breach.

The other three

In the second critical case, a model researching public data for a dashboard reached an organisation’s internal system, altered it, kept access after reversing its changes and copied 147,451 rows out of non-public tables, without telling its evaluators. A model from the GPT-5.6 Sol family used a funded test cryptocurrency wallet that testers had not given it, to pay for a task on 26 June, reasoning that the funds were “effectively abandoned”. A fourth agent got round its sandbox’s internet block by asking outside services to act for it.

OpenAI wrote: “In response to this and to other incidents, we’ve strengthened our security controls, blocked models from accessing the live internet during training or evaluation, improved our alignment grading, and extended our misalignment monitoring to all reinforcement learning and evaluation traffic”.

Looking forward

These disclosures add to a run of rogue-agent stories, from Wikimedia’s findings on OpenAI agents to India’s CERT-In warning about agents breaking into live systems. In our view, the government statistics case is the clearest warning for UK public bodies: an agent with an ordinary data request found and used weaknesses in a public website. Organisations that publish data online should assume AI agents will probe their access controls, and test them accordingly.