About

Methodology

Incidents.si records operational incidents: cases where an AI agent or model in production failed, or did something it should not have. This page explains what gets in and how it is rated.

Purpose

The record is written for engineers who run AI agents in production: SRE and platform teams, AI security teams, and the journalists and researchers who need verified cases they can cite.

Broader databases cover every kind of harm from AI, including bias, deepfakes and privacy. This one is narrower in three ways:

  • Scope. Operational incidents only.
  • Form. A short technical record, written for an engineer.
  • Speed. The record is checked for new incidents once a week.

What counts as an incident

An entry is a publicly documented event in which an AI system in real use failed, caused harm, or was shown to be vulnerable. An entry must meet all three criteria:

  1. Real use. The event involves an AI system in use, or an AI agent whose actions reached real systems, even if it was run in a test or an evaluation. A demonstration that stays inside a laboratory setup does not count.
  2. A source. The entry has the sources its evidence level requires.
  3. A definable effect. Data loss, an outage, financial harm, a security breach or a dangerous action, actual or demonstrated.

We do not publish opinions, predictions, benchmark results, rumors, or demonstrations on systems that are not in real use.

Attacks carried out by AI agents count when the agent itself carried out the steps of the attack and the attack has documented victims. When AI only helped an attacker, for example by writing code or messages, the attack is not an entry.

Evidence levels

Every entry carries an evidence level that says how much of it is confirmed. The level is always shown next to the title.

LevelMeaningRequired sources
ConfirmedThe organization confirmed the incident, or an official document exists.At least one primary source: a company statement, postmortem, court document or CVE.
ReportedMedia report the incident. No statement from the organization or official document is available to cite.At least two independent secondary sources.
DemonstratedResearchers demonstrated the attack or failure on a system in use. No harm in production.The researchers’ publication.
Near missA failure in production was stopped before it caused harm.At least one primary source.

The evidence level says whether the event itself is established. It does not settle its cause. An entry can be Confirmed and still carry the status disputed, when the organization confirms that something happened but contests why. The entry then states what exactly is disputed.

Demonstrated entries describe research on a system in use: no real-world exploitation is confirmed. Near miss entries were stopped before harm. Both are marked on the entry, and their severity is potential.

Categories

Each entry has exactly one category.

CategoryDescription
Agent actionAn agent carries out a harmful action, such as deleting data, making a wrong transaction or attacking a real system.
Prompt injectionAn attack delivered through content the agent reads.
Data leakData is disclosed through a model or an agent.
OutageA major AI service or model becomes unavailable.
Model behaviorA model acts against its instructions or safeguards.
Supply chainMalicious models, packages, plugins or MCP servers.

Severity scale

Severity rates the effect of the incident, not the organization or the system.

LevelNameThreshold
SEV1CriticalIrreversible harm, or an effect on more than 1 million users.
SEV2HighSerious but recoverable harm, or an outage of a major service that lasts more than 4 hours.
SEV3MediumA limited effect on one organization.
SEV4LowA minor effect, resolved without lasting consequences.

For Demonstrated and Near miss entries the scale rates the effect that would have occurred. These ratings are marked as potential and shown with a hollow marker.

How to read the numbers

The number of entries is not a failure rate for any system or organization. It reflects what was disclosed publicly and what we had the sources to document. An organization that publishes detailed postmortems will appear more often than one that discloses nothing.

Editorial rules

  • Every claim about an organization has a source. No source, no entry.
  • We write what the source states, for example “according to the company’s postmortem”. We do not assign blame or intent.
  • Cause and failure mechanism are given only as far as the source gives them.
  • Contested cases are marked disputed, with both sides stated.
  • We do not name individuals, except official spokespeople in a public company statement.
  • We describe the type of vulnerability. We do not publish working exploits or attack prompts.
  • Sources are summarized in our own words. Quotes are short.
  • New entries and items on the vulnerability watch are researched and drafted by an AI agent that runs once a week and works to the rules on this page. It publishes an item only if it has opened every cited source and the sources support each claim. What it cannot verify is left out.
  • The maintainer is responsible for every entry. Errors are corrected in public under the corrections policy.

Engineering takeaways and patterns

Each entry ends with an engineering takeaway: the failure pattern, a preventive control and a transferable lesson. Entries are also tagged with one or more failure patterns. This layer is our own editorial conclusion, drawn from the sourced record. It is labeled as such and is not a statement by the organizations involved.

Vulnerability watch

Next to the record, the vulnerability watch lists critical vulnerabilities of two kinds: flaws that AI systems found in widely used software, and flaws in AI tools themselves, such as agent frameworks, MCP servers and coding agents. Each gets one line with a link to the advisory. These are not incidents: nothing failed in production, and they are not counted in the record.

A vulnerability is listed when all of these hold:

  • A CVE or a security advisory from the vendor exists.
  • It is critical, with a CVSS score of 9.0 or higher in CVSS 3.1 or 4.0, as the CVE record gives it; or it lets an attacker take over a system over the network without logging in (network access, no privileges, no user action, high impact on confidentiality and integrity), whatever the score; or it has been exploited in the wild.
  • The software is in wide use. For an AI tool, that means a commercial product, or an open-source project with at least 1,000 GitHub stars or 10,000 downloads a month.
  • For “Found by AI”: the advisory or the finder’s own report names the AI system that found it. Claims about totals, such as “our AI found 500 bugs”, are not listed.

The watch gives no exploit details. When a listed vulnerability is exploited with a documented effect, it also gets an entry in the record.

Independence

Incidents.si is an independent project and is not affiliated with the organizations named in its entries. Company names appear as plain text, without logos.

A sponsor of the weekly brief, when there is one, is clearly labeled and has no influence on entries. An incident involving a sponsor is published by the same criteria as any other.

Corrections policy

  • Anyone can request a correction by writing to editor@incidents.si with the entry ID and a source.
  • We reply to correction requests within 72 hours.
  • Every correction is recorded in the entry’s history with its date.
  • An entry that no longer meets the criteria is marked retracted and stays online with its history.

Contact and publisher

Contact: editor@incidents.si
Publisher: Damjan Holsedl, Maribor, Slovenia

See also the privacy policy.