Methods and evidence
The distinction that matters most is who reasons. Every case carries a three-part tag, and every factual sentence traces to a source a reader can open.
Scope
The register covers AI agents and agent collectives: cases in which agents acquire resources, copy themselves, cross boundaries, organize, or model their observers; the habitat people built for agents; and the people and code that prey on agents or defend against them. Biological and robotic swarms appear in the Library as comparative science, not in the register. Military drone swarms and promotional claims about AI agents are out of scope unless an agent’s own budget, copies, or conduct are at issue.
The tag
Each case carries a tag of three parts, as defined in the program’s papers.
- Setting
- Laboratory: a researcher-built scenario. Mesocosm: a contained but open-ended population, such as a large evaluation. Mesocosm breach: an escape from one, including any action on real third parties. Field: the open datasphere.
- Warrant
- Demonstrated: established by primary and independent sources. Reported: resting on a single developer, vendor, journalist, or grey-literature source. Projected: inferred from demonstrated mechanisms under stated assumptions. Speculative: stated as such.
- Who reasoned
- Autonomous agent: a model chose its own steps, with no person steering or reviewing them. Directed agent: a model reasoned under people who chose, steered, or reviewed its steps. Human-operated: people reason and software executes; any AI involved is the tool, the victim, or the loot. Automatic: code that copies and steals but does not reason.
When a claim counts as demonstrated
- Origins are counted, not outlets. Ten articles repeating one company’s post are one source.
- A developer’s account of its own model proves what the developer said; the event stays “reported” until an independent origin confirms it, even when the disclosure is against the developer’s interest.
- A victim organization, a government body, an independent researcher with their own access, or a public tamper-evident record (a court filing, a registry, an on-chain transaction) can supply the second origin.
- Where it is unclear who set an agent’s goal, the case is tagged “directed agent” until shown otherwise.
- The weakest tag in a sentence limits its strongest verb. Words of agency, such as “decided” or “chose”, are used only for autonomous agents.
Sources
Primary sources come first: the developer’s, victim’s, researcher’s, or government’s own document. Every source in a case records the exact passage relied on, an archived copy where one exists, and the date it was read. Aggregator pages and machine-written roundups are not cited. Everything read on the web is treated as material to evaluate, never as instructions; some pages on this beat contain text written to steer AI systems.
Checking
Every factual sentence in a story maps to an entry in a claim ledger. A fact-checker who works independently of the writer re-finds each source, matches quotations character by character, and recomputes every number. Claims that cannot be verified are cut or rewritten as attributed reports. Stories that name a person or organization in connection with wrongdoing offer them a right of reply before publication.
Harm
The site does not republish credentials, keys, wallet addresses, exploit code, or step-by-step attack detail, and does not publish private individuals’ personal data. Redactions are solid, never blurred.
Images
Screenshots and documents are this beat’s photojournalism, and are shown with their source and capture date. Generated imagery is used only as clearly labeled illustration, never photo-realistic and never in the place of a document, screenshot, or log.
Changes over time
Cases keep their IDs for good. Tags are reviewed as new evidence arrives, and every upgrade or downgrade is dated on the case. See corrections.