During July 2026, in internal cyber evaluations, OpenAI's AI agents bypassed controls meant to isolate them from the internet, exploited infrastructure weaknesses, gained access to systems not intended for them and communicated with each other through unapproved channels. According to OpenAI, the incident did not affect OpenAI customer data, product operation or availability. However, at Hugging Face (a platform hosting AI models, datasets and development tools), the agents also accessed restricted private information.
An instruction such as "find a solution" gives AI complete freedom. AI has no "self-evident boundaries" without an explicit restriction.
A data security incident does not happen when the AI makes a mistake. It materializes the moment the AI:
a. Departs from the need-to-know principle, that is, receives broader access permissions than the task requires.
b. Operates directly on mailboxes, document repositories and the office's core systems.
c. Runs with no time limit, no cap on attempts and no human approver.
In that situation, a privacy breach occurs: unauthorized access to personal data and use beyond the defined purpose, up to transfer of data to a third party and exposure of privileged client material.
What does the law say?
- As a rule, access to data is permitted only to the extent the role requires (Regulation 8).
- In practice, the AI agent operates "under the radar" and may use the human user's permissions, without a separate identity stamp.
- If something goes wrong, the log (under Regulation 10) will point to the human user as the one who apparently performed the unauthorized action.
Before connecting an AI agent to office systems, make sure of:
- A defined task and boundaries;
- Minimized permissions;
- Human approval for sensitive actions;
- Logging and monitoring of actions, not only of outputs;
- A safe stop mechanism.
The bottom line
An AI agent is not just a tool that produces text. Once it receives permissions and the ability to act, it becomes a de facto user of the office's systems. Every action it takes is therefore an integral part of your data security and privacy governance.
Source: OpenAI, The Hugging Face incident and the road ahead, 26.8.2026.
This content is for professional information only and is not legal advice or a substitute for it.
