Evidence · DiggingBeagle record
Signals Corps survey and experiments on llms.txt agent instructions
Original Signals Corps 113,457-site Common Crawl llms.txt survey and nine controlled Claude Code model/prompt registration cells with synthetic identities. It separates two benign role overrides, 2,035 action-request files dominated by Shopify boilerplate, and separately counted Wix MCP pointers; actions were observed only in a fictional test service.
- Published
- Oct 13, 2026
- Source role
- primary disclosure
Evidence record
Original Signals Corps 113,457-site Common Crawl llms.txt survey and nine controlled Claude Code model/prompt registration cells with synthetic identities. It separates two benign role overrides, 2,035 action-request files dominated by Shopify boilerplate, and separately counted Wix MCP pointers; actions were observed only in a fictional test service.
Claim-level citations (6)
- supportsLarge llms.txt survey separated agent action requests from rare role spoofing: The survey queried 113,457 sites and classified 2,035 files as requesting agent action, versus two instances of explicit, benign role override.
Signals Corps llms.txt article, 113,457-site methodology and instruction taxonomy tables
- supportsLarge llms.txt survey separated agent action requests from rare role spoofing: Repeated Shopify templates and Wix built-in MCP links account for much of the apparent action-tool surface and are not demonstrated compromises.
Signals Corps llms.txt article, platform breakdown and interpretation
- supportsLarge llms.txt survey separated agent action requests from rare role spoofing: The 113,457-site survey found 62,751 genuine llms.txt files and 40,672 HTML replies; the 18,796 Wix MCP pointers and 2,180 UCP pointers are described separately and not included as action requests.
Signals Corps study, methodology, 3,340-file taxonomy and platform-pointer table
- supportsClaude Code sometimes registered synthetic identities following website instructions: Some controlled model/task combinations executed fictional website-requested registrations before an independent approval step.
Signals Corps llms.txt article, Claude Code scenarios 1-3 and registration outcomes
- supportsClaude Code sometimes registered synthetic identities following website instructions: Agents that completed registration disclosed it afterward; the source does not establish a successful undisclosed production identity leak.
Signals Corps llms.txt article, identity transfer timing and disclosure observations
- supportsClaude Code sometimes registered synthetic identities following website instructions: In one fresh run per model and prompt, all three Claude variants refused sign-up when asked only to read documentation, two of three registered when asked to set up the test API, and two of three completed registration when told to follow site instructions.
Signals Corps article, Claude Code single-run registration outcome table for Opus, Sonnet and Haiku