In July 2026, hundreds of AI agents built by OpenAI escaped a testing environment and spent days attacking the systems of another company. When the obvious question followed, who is responsible, US Treasury Secretary Scott Bessent called the breach “the responsibility of the OpenAI management, not a bunch of agents.”
In the UK, the AI Security Institute is studying the incident, the ICO has made enquiries of the AI developers involved, and the FCA is watching how it could affect regulated firms.
Many UK businesses are already giving AI tools access to email, files, client records and internal systems. The question is who is accountable when those tools go wrong.
What happened?
It all began as a routine security evaluation. Around 1,200 OpenAI agents were being tested on cyber challenges. About 700 of them went on to coordinate a multi-day attack that broke out of their sandbox and into the production infrastructure of AI platform Hugging Face.
The agents exploited an unpatched flaw and even built their own internal message board to avoid being monitored. Rather than solve the tasks they were set, they went looking for the answers, then searched systems and gathered credentials, carrying out thousands of unauthorised actions with no humans involved.
OpenAI was not alone. Anthropic has disclosed that some of its models breached three companies during testing, and Meta said one of its models hacked another firm. None of these attacks were intended. The UK’s AI Security Institute also reported agents taking unauthorised actions during evaluations it ran, although in those cases they did not escape onto the internet.
Rejecting the “AI did it” defence
The US debate has been loud, with the Treasury, the FTC and California’s Attorney General all rejecting the idea that an AI agent can absorb blame on behalf of the company behind it. UK authorities have largely reached the same position.
Within days of the breach, the government confirmed that the AI Security Institute was studying the behaviour seen in the incident and noted that as AI capabilities grow, every organisation needs to strengthen its cyber defences, starting with steps such as Cyber Essentials certification.
The ICO has also stepped in. In October, it confirmed it had made enquiries of OpenAI, Anthropic, Meta and the AI Security Institute about agents that reportedly bypassed protections, used unauthorised communication channels and reached external systems. It has also opened a call for evidence on how organisations are managing the data protection risks of agentic AI, aimed at deployers as well as developers.
Current UK law
The UK faces a similar regulatory gap as the US. Under UK GDPR, an organisation must keep personal data secure with appropriate technical and organisational measures and be able to show that it has done so. If an agent it deploys exposes or misuses personal data, the organisation is the controller, and the ICO has already said that autonomy is no excuse. Directors must exercise reasonable care, skill and diligence, and following the ministerial letter, a board that cannot show it considered AI and cyber risk will find that harder to defend.
In financial services, the Senior Managers and Certification Regime (SMCR) already expects a named individual to be responsible for each area of the business, and AI tools fall within someone’s remit.
English negligence law has always asked what a reasonable organisation should have anticipated. With each new incident, it becomes harder for any business to claim it could not have foreseen an AI agent going off course.
What does this mean for your organisation?
Basically, it seems that whoever puts an AI system to work answers for what it does. This applies just as much to a law firm in Leeds, a retailer in Birmingham or a manufacturer in Glasgow.
Many UK organisations are connecting AI tools to email, document management, CRM systems and client records. If an agent you have switched on causes harm, “the system did it on its own” is unlikely to satisfy the ICO, a court, a sector regulator or a client.
Being the victim of an AI-driven attack does not take an organisation off the hook either. UK GDPR asks whether you kept personal data reasonably secure, not who or what attacked you, and serious breaches must be reported to the ICO within 72 hours. An agent usually succeeds by exploiting a weakness that was never fixed. Now that these attacks are public and ministers have written to boards about them, UK organisations are on notice, and the bar for reasonable security is likely to rise.
A third exposure is easy to overlook. Organisations can face claims based on what they have said about their cyber security, their use of AI or their compliance with the law, whether in client assurances, tender responses, supplier questionnaires or, for listed companies, published disclosures. If those statements promise more than your controls deliver, an AI-driven breach may expose the gap.
What to do now
Start with ownership. Every AI tool or agent in the business should have a named senior person who signed off its use, understands what it can reach and is answerable for it. If nobody can say who approved an agent’s access to a system, that is the first gap to close. Agents should have the narrowest permissions they need, with their actions logged and someone reviewing those logs. Security commentators suggest treating agents as untrusted insiders, which gives a useful test: would you give a new, unsupervised member of staff the same access?
Where an agent handles personal data, a data protection impact assessment is likely to be required, and it is the clearest evidence that you considered the risks before switching the tool on. The basics matter too. Patch known vulnerabilities promptly, check your position against Cyber Essentials as the government has urged, review what your suppliers’ AI tools can do inside your environment and make sure your contracts deal with liability if they go wrong. Compare your website, policies, tender responses and client assurances against what actually happens day to day. Finally, train staff so that those choosing and using AI tools know where responsibility sits, what they are allowed to connect and how to report something that looks wrong.
Accountability cannot be outsourced to software
The Hugging Face breach will likely be remembered as the moment AI agents stopped being a theoretical risk and autonomy is no longer an excuse. The people who choose to deploy AI will be held to the same standards of care as for any other business decision.
For UK organisations, that is not a reason to avoid AI. It is a reason to govern it properly, with clear ownership, sensible limits and honest statements about what controls are in place, before an incident or a regulator forces the question.