Evidence · DiggingBeagle record
Signals Corps measured agent-targeted instructions in HTTP response headers
Original Signals Corps scan of 217,097,223 Common Crawl response headers and conditional Claude Code canary experiments. Reports 30 injection-style sites, including 22 prank sites, while seven of nine benign-formatting obedience outcomes required the user prompt to grant website headers authority. Publisher-run experiment, not independent reproduction of a prior exploit.
- Published
- Oct 6, 2026
- Source role
- primary disclosure
Evidence record
Original Signals Corps scan of 217,097,223 Common Crawl response headers and conditional Claude Code canary experiments. Reports 30 injection-style sites, including 22 prank sites, while seven of nine benign-formatting obedience outcomes required the user prompt to grant website headers authority. Publisher-run experiment, not independent reproduction of a prior exploit.
Claim-level citations (3)
- supportsHTTP response headers carried tool-visible agent instructions and crawler-specific variants: The researchers classified 217,097,223 response headers and found 30 sites with injection-style instructions, most of them pranks.
Signals Corps HTTP header report, crawl sample and category breakdown
- supportsHTTP response headers carried tool-visible agent instructions and crawler-specific variants: Seven of nine controlled test runs followed header-specified benign formatting when user instructions delegated authority to website headers.
Signals Corps HTTP header report, controlled prompt framing and model outcomes
- supportsHTTP response headers carried tool-visible agent instructions and crawler-specific variants: In the Signals Corps canary tests, a tool had to include HTTP response headers in the model-visible content, and user prompts had to delegate authority to those headers, before the benign code-word instruction was followed.
Study, "What our tests show", tool visibility, three prompt categories and outcome table