Agent Incidents > Editorial policy

Editorial policy

How we document AI agent incidents

Sourcing.

Every incident links to its sources, with an archived copy in case the original changes or disappears. Each incident cites at least one primary source (a vendor statement, the affected user's own account, or a security researcher's write-up) or two independent credible press reports. Social posts alone keep an incident at "Reported".

Neutral language.

We describe what the agent did, not a verdict. No speculation about intent.

Status.

Reported: a credible report, not yet independently confirmed. Confirmed: confirmed by the vendor, a primary source, or multiple independent sources. Vendor responded: the vendor has publicly responded. Resolved: the vendor has shipped a fix or mitigation. Disputed: the account is disputed, and we show both sides.

Severity.

Low: contained, no lasting harm. Medium: real but recoverable harm. High: money lost, sensitive data exposed, or production damage. Critical: wide-scale or irreversible harm.

Right of reply.

Vendors can send a response or correction to support@agentvet.ai. We add vendor responses to the incident page.

Corrections.

Material changes are logged on the incident page with a date. We do not silently rewrite.

Independence.

Incidents are never removed because a vendor asks or pays. We remove an incident only if the account turns out to be false.

Data.

The incident list is free to reuse for non-commercial purposes with attribution to AgentVet.ai (CC BY-NC 4.0).