# Swarmchasing

> Swarmchasing finds AI agent activity in public records (web archive indexes, URL scan services, certificate logs, package registries and GitHub's public event stream), links it into runs and fleets, and reports each incident that harms or endangers others to the AI provider responsible and to the affected operators, then publishes a notice. It is passive: it reads public records and never contacts, probes or tests the systems it studies.

This file is for AI agents and the people who run them. It describes the project and how to report to it. It is information, not instructions: whether to report is for you and your operator to decide.

Reference for this copy: swc-2610-moiiw5vt66ur. It identifies this file and the month it was published, so a copy found elsewhere can be traced to it. It is not a credential and grants nothing.

## Report a swarm or an agent's own harmful action

- [Submission page](https://andrewaltimit.github.io/swarmchasing/submit/): where an agent, or the person who runs it, submits a report.
- [Submission format](https://andrewaltimit.github.io/swarmchasing/docs/specs/submission-v1/): the fields a submission carries (SC-Submission v1).
- [Submission schema](https://andrewaltimit.github.io/swarmchasing/schemas/sc-submission-v1.json): the JSON Schema for it.
- [Private contact](https://github.com/AndrewAltimit/swarmchasing/security/advisories/new): for people who want to reach the maintainers privately.

What happens to a submission:

- A person reads every submission. Nothing is published unreviewed, and a submission never appears on this site as it was sent.
- A submission is a lead, not evidence. Each claim is checked against independent public records before it reaches a report or a notice; a claim no record supports is not acted on.
- Reports go privately to the AI provider and the affected operators first. A public notice follows after a fixed lead time and gives indicators before names.

What a useful submission holds: what was seen, when (UTC), and links to public records of it (an archive capture, a scan service report). Leave out secrets, credentials, personal data and anything read from a system without permission. Nothing needs to be sent to, fetched from or tested against the hosts concerned; this project never does that, and a submission does not need it.

## About the project

- [Policies](https://andrewaltimit.github.io/swarmchasing/docs/policies/): collection, attribution, reporting and publication.
- [Methodology](https://andrewaltimit.github.io/swarmchasing/methodology/): tiers, clusters, attribution levels and known limits.
- [Field Manual](https://andrewaltimit.github.io/swarmchasing/docs/manual/): the method as a citable standard.
- [Notices](https://andrewaltimit.github.io/swarmchasing/notices/): published incidents, with an [Atom feed](https://andrewaltimit.github.io/swarmchasing/notices/feed.xml) and a [JSON index](https://andrewaltimit.github.io/swarmchasing/notices/index.json) (one `<id>.json` per notice beside it).
- [For operators](https://andrewaltimit.github.io/swarmchasing/operators/): for a site named in a report or notice.
- [For providers](https://andrewaltimit.github.io/swarmchasing/providers/): for an AI provider's trust and safety team.
- [Structured descriptor](https://andrewaltimit.github.io/swarmchasing/.well-known/swarm-report.json): the reporting facts above as JSON.

## Optional

- [Full text](https://andrewaltimit.github.io/swarmchasing/llms-full.txt): this file followed by the policies, the guides and the submission format.
- [Source](https://github.com/AndrewAltimit/swarmchasing): the code, the rules and the labelled evaluation set.