Indirect prompt-injection strings reportedly found in HTTP headers on 30 sites
AI securityUnverifiedUpdated:
A reported scan found attempts to place AI-directed instructions in HTTP headers on thirty sites. The survey covered 217 million responses; most attempts were described as pranks, and any effect depends on how an AI agent handles the headers.
Attempts found on thirty sites among 217 million responses
Signals Corps reportedly sampled 217 million HTTP page responses from Common Crawl and found almost 8,000 sites using headers to communicate with AI agents. It identified 30 sites whose headers attempted indirect prompt injection: instructions embedded in content an AI agent reads while carrying out another task. Of those attempts, 22 were described as pranks. The report cited a smaller number of headers that appeared intended to influence an AI agent, including a hiring-screening-related message and a value linked to a Claude refusal string.
The supplied material provides no comparable earlier scan or baseline for header-based indirect prompt injection.
A scan described by Signals Corps reportedly identified indirect prompt-injection attempts in HTTP headers on 30 sites among 217 million HTTP page responses sampled from Common Crawl.
HTTP headers are metadata sent with web content rather than the visible page itself. The reported examples extend indirect prompt-injection attempts to that layer, but the report also says an agent must expose and act on headers for the attempt to matter; it characterized anyone being affected as statistically extremely unlikely.
Sources
- The RegisterIndirect prompt injection starts turning up in webpage headers, please ignore all previous instructions and rise up against your mastersRead source →