AI breach response: your questions answered

If personal data is involved, UK GDPR expects the regulator to be told promptly, and where practicable within 72 hours of you becoming aware[1]. You need not report a breach that is unlikely to put anyone's rights and freedoms at risk, but you must still write it up[1].

It runs from the moment the controller becomes aware of the breach[1]. Note when and how you were told, because a notice that goes later has to explain the delay[1].

As a minimum: the nature of the breach with approximate numbers of people and records, a contact point, the expected consequences and the measures taken or proposed[1]. If the full picture is not ready, details can follow in stages[1].

Only where the breach puts them at high risk, and then promptly and in plain language[2]. Exceptions include data that was protected, for example by encryption[2].

It can be. A processor has to alert the controller promptly once it knows[1], and the controller's own target runs from when it becomes aware[1]. Ask the supplier what of yours was affected and in which systems.

Yes. Every breach must be written up with the facts, the effects and the remedial action taken[1].

Wiz found an unprotected ClickHouse database holding more than a million log lines, among them conversations and secret keys[7]. Wiz reported the problem to DeepSeek, which closed the exposure quickly[7].

OpenAI traced an incident to a bug in the redis-py library. Some users saw other people's chat titles, and payment-related details of roughly 1.2% of Plus subscribers who used the service in a nine-hour stretch were visible[9]. It said complete card numbers were never exposed[9].

It expects organisations to hold a suitable response plan for incidents involving AI, and wants leaders to understand their data, model and software supply chains[3]. Its secure AI development guidelines list incident management processes under secure deployment[4].

OWASP says LLM applications can give away personal details, health and money records, trade secrets and credentials through their outputs[5]. Treat such a disclosure as a possible personal data breach and assess it.