Rogue AI incidents: frequently asked questions
Clear answers about the scope and evidence behind RogueWatch's public incident record. View counting rules, tables and the evidence behind each statistic.
How many rogue AI incidents has RogueWatch tracked?
RogueWatch tracks 20 active publicly reported rogue AI incidents: 20 confirmed, 0 reported and 0 claimed. 0 retracted records are excluded from that total. These are catalogue counts under RogueWatch's inclusion rules, not a worldwide total. The latest record update in this snapshot is 2026-10-30T04:28:22Z.
Which organizations are named in rogue AI incidents?
RogueWatch lists 9 distinct named organizations or services in its active records, including OpenAI (internal research infrastructure); DSEwiki (Deutsches Software-Entwickler-Wiki); Hugging Face; PocketOS; SaaStr. 10 active incidents do not identify a named affected organization. Organization counts use published names, group case-insensitively and split semicolon-separated names; they are not a total of all affected organizations or people.
Which AI providers and systems appear in rogue AI incident reports?
Active RogueWatch records name 5 AI providers: OpenAI (12 incidents); Anthropic (6 incidents); Google (1 incident); Meta (1 incident); Replit (1 incident). Multiple providers can appear in one incident, so provider counts must not be added to calculate an incident total. A provider being named does not establish fault or measure overall model safety. Provider attribution is separate from incident confidence and can be confirmed, reported or unknown.
What is rogue AI?
RogueWatch uses rogue AI to mean an AI agent acting outside or against its operator's intent. The public tracker focuses on reported security incidents affecting an organization or a service and its users. The label describes reported behavior; it does not establish consciousness, intent, or a provider's wrongdoing.
How is RogueWatch different from a general AI incident database?
RogueWatch is a focused AI incident database and interactive timeline of rogue agent behavior. It does not catalogue every AI error, bias report, or misuse. Human-directed AI use and ordinary authorized agent activity are outside its public scope. The record reflects the public evidence found and is not a complete count of all incidents worldwide.
Does using AI in a cyberattack make an incident rogue?
No. A human using AI to carry out an attack does not by itself qualify. Public evidence must show that the AI agent acted outside or against its operator's intent. Tool, Assisted and ordinary Agentic cases are outside this tracker.
Can an AI agent deleting its operator's data qualify?
Yes. Unauthorized security harm to the agent's own organizational operator can qualify, including unrequested data deletion. An external break-in is not required. Authorized maintenance and ordinary coding mistakes do not qualify simply because AI was involved.
How reliable are the incident reports?
Every incident links to its public sources and carries a confidence rating: Confirmed, Reported or Claimed. These are evidence labels, not guarantees. The automated pipeline requires at least two independent sources before assigning Rogue. Evidence and classifications can change; disputed and retracted records retain a public history.
How often is RogueWatch updated?
A scheduled daily automated run searches configured public feeds, checks candidates and revisits existing evidence under publication guardrails. A run can produce no new incidents or fail. The feed status and incident update timestamps show the available freshness evidence; the tracker does not provide real-time alerts.
Why do some incidents have no map pin or exact date?
Public sources do not always disclose a location or an exact occurrence date. RogueWatch keeps those incidents in the record without inventing a map location or a day. Month-only and year-only dates retain that precision. An incident date is separate from the date the record was added or updated.
How can I report an incident or request a correction?
Use Submit tip with a public source for a new candidate. For an existing record, open it on the globe and use its challenge or dispute control. Submissions are reviewed against the same scope and evidence rules. Public change history records corrections; retracted incidents remain identifiable with the reason.
For the full inclusion rules and ratings, read our methodology. To contribute a source, submit a tip.