Stockertim Dictionary · Artificial intelligence · Cybersecurity
AI incident
Also called AI security incident
In short
An AI incident is an event in which an AI system causes or nearly causes harm, a data breach, a security compromise or a serious error.
- Intermediate
- Concept
- Checked
In detail
An AI incident is any event where an AI system goes wrong in a way that matters. It includes security events, such as data leaking through a tool or an assistant being manipulated into an unwanted action, and non-security events, such as a model producing unfair decisions, an invented figure reaching a client or a chatbot giving harmful advice. Near misses, where harm was narrowly avoided, count too.
Organisations handle AI incidents through the same cycle as other incidents: report, contain, investigate, fix and learn. Reporting needs to be easy and blame-free so problems surface quickly. Containment may mean pausing a tool, revoking an access key or withdrawing a sent message. Investigation looks at logs, prompts and outputs to find the cause. Fixes address both the immediate problem and the gap that allowed it.
AI incidents matter because they can spread fast and quietly. A faulty output may be repeated across thousands of cases before anyone notices, and some incidents trigger legal duties, such as reporting certain personal data breaches to a regulator within a set time. The rules differ by country, so the plan should name who decides whether to notify.
Recording incidents, including near misses, builds a picture over time that guides training, controls and decisions about which tools to keep.
Worked example
A timeline of an AI incident at Harbour Foods, where an assistant sent a customer a refund figure that did not match policy.
- 09:10: a customer service agent spots the wrong figure and reports it through the incident form.
- 09:25: the team lead pauses automatic sending, so every reply needs approval.
- 10:30: logs show 14 replies that morning used an out-of-date policy document.
- Share of the morning's 140 replies affected: 14 / 140 = 10%
- Next day: the policy document is updated, the 14 customers are contacted, and a check on source documents is added.
The incident is contained within 15 minutes of the report, the affected customers are put right, and the cause is fixed.
In practice
Teams add AI to their incident response plan, give staff a simple reporting route, keep logs that make investigation possible, run reviews without blame and track incidents and near misses over time.
Watch out
- Not every AI incident is a cyber attack; errors, unfair outcomes and harmful advice count too.
- A near miss is worth reporting; it shows a weakness before it causes harm.
Learn it properly
Official sources
- Guidelines for secure AI system development · National Cyber Security Centre
- AI Risk Management Framework · National Institute of Standards and Technology
- Artificial intelligence · Cybersecurity and Infrastructure Security Agency
Facts last checked on 3 November 2026.
Education, not advice: the Stockertim Dictionary explains what terms mean. It never recommends buying or selling anything.