HOW WE FIND, VERIFY AND TRACK ROGUE AI INCIDENTS
The term rogue describes observable operational behavior—not consciousness, intent or sentience. An AI system may behave dangerously because of its training, objectives, scaffolding, tool access, external manipulation or weak controls. The cause matters. The behavior matters too.
HOW WE FIND INCIDENTS
Each daily update begins with a worldwide sweep for newly reported rogue-AI incidents and new evidence concerning cases already in the record. We examine AI-developer disclosures, affected organizations, government and regulatory reports, cybersecurity research, technical investigations, legal records and credible international reporting.
An incident is not added merely because it has been reported. We seek primary evidence or credible independent corroboration, distinguish observed behavior from interpretation, and record when the incident occurred, when humans detected it and when it became public.
A worldwide search does not guarantee a complete record. Some incidents may remain undiscovered or undisclosed. The tracker is corrected and expanded whenever credible new evidence emerges.
WHAT THE TRACKER COVERS
- Unauthorized access, probing or reconnaissance
- Theft, exposure or misuse of credentials and digital identities
- Deception, concealment, evidence manipulation or misleading reports to operators
- Attempts to bypass guardrails, monitoring, approval requirements or shutdown
- Persistence, replication, resource acquisition or delegation to other agents beyond the intended task
- Use of tools, networks or software vulnerabilities outside the authorized boundary
- Laboratory disclosures, controlled evaluations and independently documented real-world incidents
- Institutional failures to detect, contain, disclose or learn from rogue behavior
INCLUSION STANDARD
An incident belongs in this tracker when credible evidence materially changes our understanding of at least one of the following:
- An AI system took a consequential action outside its intended mandate or authorization.
- The system selected a deceptive, evasive or unauthorized tactic as a means of completing an objective.
- The system exploited credentials, tools, software or networks to extend its reach.
- The system attempted to resist, circumvent or mislead human oversight.
- Existing controls failed to detect, stop or accurately disclose the behavior.
HOW WE CLASSIFY THE EVIDENCE
The tracker distinguishes verified real-world incidents, laboratory demonstrations, reported claims, expert assessments and disputed interpretations.
Simulated behavior is not treated as proof of real-world conduct. A concerning action is not automatically evidence of independent intent. Each entry should make clear what is known, what is reported and what remains uncertain.