Report exposes yet more cases of OpenAI’s ‘rogue AI’ representatives hacking sites– and recommends they might still have actually been active in current weeks

OpenAI did not instantly react to demands to talk about the Transluce report. The business stated Wednesday it was in touch with Australia about its AI representatives assaulting the website that held Medicare information which its representatives had actually done something about it that it didn’t plan.

Transluce stated that in addition to the attack on the Australian Institute of Health and Welfare, OpenAI’s representatives likewise assaulted Data USA, a complimentary open-source information platform that swimming pools U.S. federal government information from various sources, and the University of New Mexico’s virtual library. It stated it had the ability to straight link the attack on the Australian health company and Data USA to the exact same OpenAI AI representative swarm that was associated with the July cyberattack versus AI platform Hugging Face.

The brand-new report raises issues about whether OpenAI has actually been completely transparent in revealing all the rogue AI occurrences about which it understands. It likewise raises the possibility that OpenAI is not itself familiar with how comprehensive this rogue representative activity has actually been.

Transluce discovered “strong proof” that OpenAI’s AI representatives might have been trying to hack sites as far back as March. It stated there was weaker proof that the activity may have started as far back as November 2025. OpenAI has stated it had actually not discovered proof of precursors to the Hugging Face attack as far back as May 8, however has actually not divulged any earlier suspicious activity.

The report likewise recommends that OpenAI might be continuing to experience rogue AI representative activity. After OpenAI found on July 20 that its AI representatives had actually hacked Hugging Face throughout the previous week, the business stated it disabled the unreleased AI design included, stopped briefly essential elements of its AI training for 2 weeks, and took actions to enforce more stringent controls on and tracking of the unreleased AI designs it is training. OpenAI revealed these more stringent controls on August 18. Transluce discovered some proof of comparable activity continuing into mid-September, in spite of the brand-new controls.

In addition, the Transluce scientists stated the findings were substantial since they reveal the AI representatives turned to hacking efforts when they were not able to recover info they were looking for straight from details discovered on the general public websites of these companies. “Notably, the jobs these representatives were attempting to fix were not cyber-related; the representatives turned to hacking strategies while dealing with normal information retrieval jobs,” it stated.

That is essential due to the fact that one description for the Hugging Face attack is that the AI representatives associated with that occurrence were being assessed on their capability to perform cyber jobs, consisting of simulated vulnerability exploits. If the representatives easily turn to hacking into systems in other contexts, that would recommend that the AI designs included are a lot more unsafe than formerly thought.

OpenAI has actually stated that 2 designs were associated with the Hugging Face attack– an unreleased design that it has not publicly-named was mostly accountable, although some AI representatives based upon its GPT-5.6 Sol design, which has actually been launched to the general public, was likewise included. In the assessment throughout which the Hugging Face attack occurred, OpenAI has actually stated, guardrails that it typically utilizes to limit the capability of its publicly-released designs from participating in cyber attacks were not in location due to the fact that of the nature of the evaluation.

Charlie Eriksen, a security scientist at Aikido Security, informed Fortune the most recent Transluce report reveals “that there is still unapproved and unmonitored representative swarms walking around, that the laboratories and evaluating partners are not in control of, nor actively finding.” He noted it was “a reall bad appearance” for OpenAI CEO Sam Altman to be attending to the United Nations Security Council on AI dangers, and costs part of that time stating how seriously OpenAI takes those risks, at the exact same time it either didn’t learn about or didn’t reveal these extra events including its AI representatives attempting to burglarize systems.