Methodology

Source and scope

This site's current sole source is Anthropic's threat intelligence report Detecting and countering misuse of AI: September 2026, released on September 10, 2026 (154 pages, English; main observation window December 2025 to August 2026).

This site does not investigate cases on its own, nor does it incorporate rumors from other sources into cases. All behavioral descriptions, scale figures, and attribution judgments come from this report; editorial protection recommendations are labeled separately from report content.

Splitting criteria for the 42 units

42 is the number of article units this site split for readability, not 42 separate cases. The original report is organized into seven harm chapters; this site further splits within chapters by heading position to facilitate case-by-case search and reading.

A few splitting examples: GTG-50021 in the cyberattacks chapter originally appeared within the AI supply chain topic; this site elevated it to a standalone unit for search and popularization. Biological cases 4 and 5 were narrated in the same section in the original; this site keeps them as one unit with two source IDs coexisting. GTG-16012 and GTG-16003 were narrated together in the same section in the original; this site combines them into one article, with the two subjects expressed in separate paragraphs internally.

Conversely, one GTG unit may itself contain multiple accounts, organizations, or actions. GTG numbers are internal tracking identifiers used by Anthropic, not judicial case numbers; article counts, GTG counts, account counts, and action counts are different counting objects and cannot be converted into each other.

Three types of labels: facts, assessments, recommendations

This site divides content into three categories and presents them separately: report facts (events and figures directly described by the report), report assessments (the publisher's attribution and confidence judgments, such as 'medium confidence'), and editorial recommendations (protection tips added by this site).

Report assessments are always presented in the wording of 'the report states', 'Anthropic attributes ... to ...', not written as this site's independent findings; the three categories of information are not merged into any form of 'credibility score'.

Numbers and wording scope

Numbers preserve the original report's comparison operators, time windows, and object scope: 'more than 4,700' is not written as '4,700', 'at least 25,000 unique interaction targets' is not written as '25,000 victims'. No unsupported calculations are made.

Negative evidence must be preserved: where the report says 'X did not happen', the page states it truthfully. Unknown items are marked as 'not provided in the original report', not filled with templated guesses. Attribution wording is not upgraded: 'medium confidence' is not written as 'confirmed', and allegations involving companies are not written as conclusions of illegality.

Translation process and review

Translation is not sentence-by-sentence matching: first build a fact package from the English original report, registering numbers, time windows, attribution, and negative evidence item by item; then write the editorial master draft; finally paraphrase into Chinese. Key facts are aligned with the original item by item; sentence organization can be rearranged for Chinese readers, but the factual scope remains unchanged.

Native-language editorial review and (required for high-risk topics such as biology) professional review are recorded separately; content that has not completed review does not display a 'reviewed' mark. After errors are found, the revision history distinguishes between 'source correction' (original report update) and 'translation/editorial revision' (this site's changes), published at the bottom of the page.