What Followed the Login?
Spain’s privacy regulator says its first AI-agent breach notification by the affected organisation remains under analysis, proving neither autonomy nor a world first. Reportedly, generic-file vulnerability search preceded successful login, then application search, personal-data changes and invoice access.
The organisation, actor, model, vulnerability, duration and affected-person count remain undisclosed. Model or provider-infrastructure compromise and malicious model design are unproved. No trend follows. Agents can plan, use tools, execute code, consult sources and adapt, potentially shortening response; incident speed was unmeasured.
- Map ordinary and service accounts, applications, records and operations; test authorisation beyond authentication.
- Control consequential payments, exports, sensitive records and permissions.
- Exercise and measure suspension, session/token revocation plus investigation handover.
- Log reads, exports, changes and failures with context; minimise personal data.
- Assign abnormal-behaviour alerts an owner and response action.
- Report confirmed access and changes; keep actor, tooling and autonomy provisional.
BlackTree guidance, not findings or prevention guarantees.
Privacy and legal owners should assess actual local processing and consequences, not a technology label.
Stolen-session analysis covers another identity boundary; no case link established.
AEPD, 14 September 2026; no public clock.
