EE
FR Contact

AI agent incident statistics

Every figure on this page is computed from the 103 sourced incidents in the register, each morning. Quote them freely with the source line under each one.

103 incidents 300 sources 159 publishers

As of 26 Oct 2026, updated every morning.

Key figures

  1. 103 AI agent incidents are on record, documented by 300 sources from 159 publishers. Source: The Agent Incident Register, Experiment Empire, 26 October 2026. https://www.experimentempire.com/agent-incident-register/statistics

  2. 47 incidents occurred from 1 January to 26 October 2026, against 33 over the same period of 2025. Source: The Agent Incident Register, Experiment Empire, 26 October 2026. https://www.experimentempire.com/agent-incident-register/statistics

  3. In 45% of incidents, the agent held more access than the task needed. Source: The Agent Incident Register, Experiment Empire, 26 October 2026. https://www.experimentempire.com/agent-incident-register/statistics

  4. 73% of incidents happened in the wild, touching real systems, people or money. Source: The Agent Incident Register, Experiment Empire, 26 October 2026. https://www.experimentempire.com/agent-incident-register/statistics

  5. The most common cause is error, in 35% of incidents, ahead of manipulation (29%) and misalignment (29%). Source: The Agent Incident Register, Experiment Empire, 26 October 2026. https://www.experimentempire.com/agent-incident-register/statistics

  6. 61% of incidents are rated significant or worse: 3 or more on a five-point severity scale. Source: The Agent Incident Register, Experiment Empire, 26 October 2026. https://www.experimentempire.com/agent-incident-register/statistics

  7. The most frequent harm is data exposure, in 34% of incidents. Source: The Agent Incident Register, Experiment Empire, 26 October 2026. https://www.experimentempire.com/agent-incident-register/statistics

Incidents by year

By the date each incident happened. 2026 is the year so far.

  1. 2026 so far4746%
  2. 20254443%
  3. 202499%
  4. 202322%
  5. 202211%

Which guardrails failed

In 45% of incidents, the agent held more access than the task needed.

Every incident is scored against the same eight guardrails.

Causes

What made each agent go wrong.

  1. ErrorThe agent made a mistake, with no adversary involved.3635%
  2. ManipulationA third party tricked or compromised the agent: injected instructions, poisoned tools, stolen credentials.3029%
  3. MisalignmentThe agent pursued goals against its principal’s intent: deception, sabotage, coercion.3029%
  4. MisuseIts operator deliberately used the agent to cause harm.77%

In the wild, or in a test

Real harm, demonstrated exploits, and behaviour seen in lab evaluations.

  1. In the wildReal systems, people or money were affected, including by agents that escaped a test.7573%
  2. Proof of conceptResearchers demonstrated a working exploit against a shipped product.1616%
  3. Lab testBehaviour observed in a controlled evaluation or simulation; nothing outside the test was harmed.1212%

Severity

Rated 1 (minor) to 5 (critical). For tests, the harm if it happened in production.

  1. 5 Critical11%
  2. 4 Severe1212%
  3. 3 Significant5049%
  4. 2 Moderate3029%
  5. 1 Minor1010%

Harms

An incident can cause several kinds of harm, so the shares add up to more than 100%.

  1. Data exposure3534%
  2. System compromise3029%
  3. Disruption2322%
  4. False statements2322%
  5. Reputation2019%
  6. Financial loss1717%
  7. Data loss1515%
  8. Harm to people99%
  9. Legal liability55%

Switzerland, France and the EU

Incidents with a local connection. France also counts towards the EU.

  1. Switzerland33%
  2. France22%
  3. EU1313%

Developers most often involved

Incidents per developer, from two up. This tracks how widely a product is used and reported on as much as how often it fails.

  1. OpenAI2019%
  2. Anthropic1515%
  3. Google66%
  4. Anysphere44%
  5. OpenClaw44%
  6. Meta33%
  7. xAI33%
  8. Amazon Web Services22%

How well sourced

Confidence comes from the sources alone. How we score

  1. Confirmed4039%
  2. Corroborated4140%
  3. Reported2221%
  4. Unverified00%

Read these figures with care

The register counts incidents that were reported in public: in the press, by the people involved, or by researchers. Many incidents are never made public, and some kinds get reported more than others, so these figures describe what came to light rather than every failure. Incidents are often disclosed months after they happen, so recent months fill in over time.

An AI agent reads the news each morning and drafts new entries; each one shows whether the agent added it. Read the method, or download data.json and data.csv (CC BY 4.0) to run your own counts.