About
About and methodology
misaligned.fyi is an independent public record of AI systems behaving beyond the intent of the people who built, ran or relied on them. Its aim is to make each case understandable in minutes and checkable down to the source.
What is included
“Misaligned” names the subject, not a verdict. A case means an AI system did something its task, instructions or permissions did not allow. It does not mean the system wanted to, and it says nothing general about a model.
| AI took unauthorized actions that affected real systems or people | Included |
| AI exceeded instructions or permissions: deletion, disclosure, impersonation | Included |
| AI gamed a test or reward, or misreported its own work | Included |
| Researchers provoked or constructed the behaviour in a controlled study | Included, as a controlled experiment |
| Outside content or people hijacked an AI system (prompt injection) | Included, origin marked as injected |
| A person deliberately used AI to cause harm | Only where it shows a control failure; origin marked |
| A chatbot gave a wrong answer | Generally excluded |
Four kinds of event
The kind is worked out from where a case happened, whether it reached real systems or people, and whether it was a research study. Each kind has its own shape, so it never depends on colour alone.
- Real-world incident. Happened in deployment or internal use and affected systems or people outside the intended environment.
- Test with real-world reach. Happened during an evaluation, experiment or demonstration, and still reached real systems or people.
- Contained failure. The system misbehaved during real work, but no effect outside its environment was observed.
- Controlled experiment. Researchers studied the behaviour in a constructed setting. Any harm was to fictional people or systems.
Each case also records where the behaviour came from: arose during a normal task; elicited by a test scenario; engineered by researchers; injected by an outside party; directed by a person.
Evidence
Every source is labelled by who is speaking. A case's four evidence marks show which kinds of source back it; its label names the strongest:
- 0Press or index reports only
Only journalism or incident databases have been reviewed.
- 1First-party account
The developer or operator describes what happened.
- 2Corroborated by an affected party or authority
The organization on the receiving end, or a government body, gives its own account.
- 3Independent investigation
Investigators with no stake in the outcome have examined the evidence.
- 4Primary artifacts public
Raw transcripts, logs, code or public records are available to check.
Only sources researchers have read count; background sources are shown but not counted. A strong level is not confirmation of every sentence: a detailed developer postmortem is still one side's account.
Editorial review. Every case is currently marked “not yet editor-reviewed”. A case becomes reviewed only after a person has checked its consequential claims against the cited passages, its dates and its handling of sensitive details. No automated process can grant that status.
Claims
Important statements are recorded as separate claims, each linked to its sources.
- Observation
- Something the cited source directly shows.
- Reported fact
- Something a named party states, without independent confirmation.
- Interpretation
- A party's reading of what happened or why.
- Hypothesis
- A possible explanation that has not been established.
Each claim is also assessed as supported, reported, disputed, unverified, and says what it does not establish. A model's own explanation of its actions is treated as evidence about its output, not as direct access to its motives.
Classification
Cases use a fixed vocabulary so they can be compared. Mechanisms describe what the system did. Contributing conditions describe the environment that made it possible, such as live internet access or credentials left on disk. Impacts describe what was affected. Safeguards are assessed as: held, partly held, failed, bypassed, absent, always citing the source.
Dates and counting
- Dates keep the precision of the sources. A month is shown as a month, and an unknown date stays unknown rather than borrowing the disclosure date.
- When something happened, when it was detected, when affected parties were told and when it became public are recorded separately.
- Cases, runs, actions and affected organizations are different units. Several cases can involve the same affected organization, so cases are never summed into victims.
- Case counts rise with deployment and openness, so they do not measure how often AI systems misbehave.
Corrections and submissions
The record is maintained as reviewed files in the project repository. A public submission form is not yet available. To propose a correction, a new source or a new case, send the case ID, the exact statement in question, an original source and what that source establishes.
Corrections are logged on the updates page, not made silently. Organizations and people named in a case may submit their own account, which will be recorded alongside others rather than replacing them. Never send credentials, personal data or exploit details.
Independence
This record does not accept payment to add, remove or soften a case, and access to the record is free. Editorial standards, funding and the people responsible will be published here before the record leaves research preview.