# AI-attack public source-collection report

Collection date: 2026-08-28

The private research archive contains **116 retrieved source records**, exceeding the requested 100-source minimum. A content/format audit found five confirmed companion-format pairs, so the set represents **111 unique publications** across 116 URL/file records. The public package retains canonical URLs, source identifiers, hashes, token counts, and duplicate decisions, but does not redistribute the downloaded third-party documents, copied evidence excerpts, or local archive paths.

## Composition

| Archive label | Count | Intended use |
| --- | ---: | --- |
| Operational CTI | 37 | Actor, campaign, malware, abuse, or incident observations |
| Vendor threat report | 37 | CTI, incident-response, or threat-landscape reporting |
| Provider/government report | 18 | AI-misuse and abuse-disruption reporting |
| Government/law-enforcement | 8 | Advisories, fraud alerts, and regional threat reporting |
| Empirical/academic | 4 | Controlled studies or measured datasets |
| Forecast/assessment | 6 | Forecasts, red-team studies, and future-risk assessments |
| Other | 6 | Requires manual relevance review before use |

Retrieval coverage is 99 direct publisher responses and 17 Jina text-reader proxy copies. Proxy-derived records must be checked against the canonical publisher page before quoting or promoting a claim; three thin JavaScript-shell responses required a selected reader representation during private collection.

The public [`source-uniqueness-report.tsv`](source-uniqueness-report.tsv) records exact hashes and duplicate-group decisions. Five additional near-duplicate candidates were retained because they are separate reports with overlapping boilerplate or publisher summaries rather than the same publication.

## Statistical-use rules

This is a purposive publication corpus, not a 116-incident prevalence dataset. A single incident may be described by several sources, and many reports are assessments, controlled studies, or forecasts rather than confirmed in-the-wild attacks. Before calculating incident rates, create an incident/evidence-cluster key, record the observation window and publication date separately, and retain numerator/denominator definitions. Treat the published source-type and relevance labels as triage hints, not as automatic truth.

The public normalized layer maps 116 source records to 111 publication entities, indexes 108 usable references, and separates the 103-publication primary denominator from context-only, non-AI, and retrieval-error material. Independent reviewers can retrieve the canonical publisher resources from [`publications.csv`](publications.csv) or the [References library](/references/) and compare them with the normalized candidates.