Something went wrong. What now?
Report it through our contact page with the product, the time and the operation id or finding reference. Meanwhile you keep your controls: hold delivery of replies, or revoke an agent's key.
What counts as an AI incident
An output or a control failure in Cyberdroid AI Detection or Cyberdroid AI Investigation that misled a decision, or could have. Following the OECD, an incident is harm that happened and a hazard is harm that could plausibly happen. Report both. Examples:
- A finding whose evidence does not support its score, or telemetry reported healthy when it was not.
- A case that claims more than its evidence references support, or treats a dead feed as proof that nothing happened.
- A request that reached an agent without an administrator's approval.
- A reply delivered to a recipient your policy did not clear.
- A translated request that changed an identifier or its meaning.
- A detector alerting in production without a recorded promotion.
- Any sign that text inside a log, alert or email changed what the system did.
What to include
- The product and version.
- The date and time, with time zone.
- The operation id and reason, or the finding or case reference.
- What happened, and what you expected.
- Evidence, redacted as your policy requires, and what you have done so far.
What you can do now
- Hold delivery. Approval and delivery are separate authorities, so you can stop replies without relying on the model.
- Revoke a key. Each agent has its own key, and revoking it stops that agent.
- Keep the record. Requests, refusals, promotions and rollbacks are already recorded. Avoid changes that would affect the evidence before you report.
