Swarmchasing Field Manual (SC-FM 1.1)

On this page

How to cite
Requirement words
Tool support
Chapters
Relation to the rest of the docs

Swarmchasing is the practice of finding AI agent activity in public records, linking it into runs, fleets and swarms, attributing it with stated confidence, and disclosing it when it harms or endangers others: to the AI provider and the affected operators first, then to the public, indicators after a short fixed lead time and the names of affected parties after a few working days.

This manual is the method as a standard. It is written so that any group can apply it with its own tools, and cite it when it does. The methodology page describes how this repository's toolkit implements it, setting by setting.

How to cite#

Cite the version and the section id:

SC-FM 1.1 §FM-3.2

Every section heading starts with its id (FM-<chapter>.<section>). Ids are stable within a major version:

an id is never reused for a different meaning;
a section whose meaning changes gets a new id, and the old one stays as a stub that points to it;
a minor version (1.1) adds sections and clarifies wording; a major version (2.0) may change requirements.

Tactic ids (SC-T1 to SC-T9, FM-4.3) and attribution labels (FM-7.7) are stable in the same way.

Requirement words#

Must, must not, should, should not and may, written in bold, carry the meanings that RFC 2119 and RFC 8174 give their capitalised forms. The same words in plain text are ordinary English.

A group that applies this manual says which sections it follows and where it departs. A report that cites a section and departs from it says so in the report.

Tool support#

The manual describes the method as a standard. Where this repository's toolkit does not yet implement a part, the section marks it (planned) and says what the toolkit does instead. The method still applies: a practitioner does that step by hand and records it in the case.

Formats other groups can implement, such as the rules file and the incident document, are specified in specs, each with a JSON Schema.

Chapters#

Chapter	Covers
1. Scope and definitions	Agent, record, run, cluster, fleet, swarm, incident, case, notice; evidence tiers.
2. Principles	The eight rules every step follows.
3. Sources and collection	What counts as a public record, the passivity test, source rationale, source biases.
4. Indicators	Rule anatomy, the SC-T tactics, artifact families, ATLAS and ATT&CK cross-walk, rule-writing standard, interaction signals, independence groups.
5. Linking	Hard links, soft links, candidate fleets, merge and split, escalation sequences, reruns, interaction edges, the cluster lifecycle, scaffold fingerprints, coordinated accounts on platforms, co-targets, weighing soft links, fleets across windows, ordered co-targets, scaffold lineage, prompt echo, co-groups, identifier generators, playbook alignment, campaigns.
6. Timing and concurrency	Timestamp kinds, cadence, bursts, rhythm, clock offset, stale dates, shared pauses, co-activity, onset groups.
7. Attribution	Four attribution layers, the evidence ladder, claim tags, labels, the naming gate, false flags, agent likelihood, signed agents, competing hypotheses, published address ranges.
8. Confidence language	The words-to-probability table every report uses.
9. Evaluation	Evaluation sets, metrics, agreement, what to publish with a number, benchmarks, agent-grade evaluation, calibrated public language.
10. Disclosure	Parties, two-stage publication and its lead times, relation to ISO/IEC and FIRST guidance, reply, consent, self-review and second review.
11. Notice standard	Notice structure, status lifecycle, withholding, lint, corrections.
12. Interoperability	STIX 2.1 profile, the `x-sc-agent-run` extension, AI incident databases, the OWASP, ATLAS, STIX, MISP and OECD crosswalk, regulatory reports.
13. Case study template	The template, a worked example on synthetic records, and three published incidents read with it.
14. Known limits and traps	Conclusions that look sound and are not.
15. Misalignment and scope	Payload, agent, scope and coordination as four findings; task and target classes; the scope tests.
16. Relation to other methods	Published investigations, provider reports, analytic frameworks and standards work, compared.
Glossary	Every defined term, with the section that defines it.
References	Every source the manual cites.

Relation to the rest of the docs#

Doc	Role
Policies	This project's policies. The manual's principles (chapter 2) generalise them.
Methodology	The toolkit's implementation: settings, thresholds, code.
Admin workflow	The admin's day with this toolkit.
For operators	For an operator who received a report.
For researchers	For a researcher applying, citing or checking the method.
Specs	Formats and schemas for exchanging rules and incidents, and how another group adopts the method.

Source: docs/manual/README.md