Prompt Injection Detected in HTTP Headers
agents
| Source: Mastodon | Original article
Researchers scanned 217 million web page responses and found 30 sites embedding hidden prompts in HTTP headers, including attempts to run code and sway AI hiring screens.
A scan of 217 million web‑page responses uncovered a new vector for prompt injection: hidden instructions embedded in HTTP response headers. The analysis, released by the Signals Corps, identified 7,959 sites that include AI‑directed cues in their headers, and 30 of those sites actually attempted to inject malicious prompts. Most of the attempts appear to be jokes, but a few reveal the potential for real‑world abuse. One header asked an AI agent to execute an npx package as a “fix”, while two others tried to influence AI‑driven hiring screens with messages such as “applicant status: 94 % match confidence”.
Why it matters is that AI agents increasingly parse not only page content but also metadata, including HTTP headers, to gather context. When attackers can slip commands into these headers, they can steer agents to run code, disclose confidential information, or manipulate decision‑making processes. The risk extends beyond pranksters; the npx example shows how a seemingly innocuous directive could trigger code execution on systems that automatically trust AI output, while the hiring‑screen manipulation hints at bias‑oriented attacks on automated recruitment pipelines.
The findings build on our earlier coverage of the same Common Crawl dataset on 15 October, where we first reported the scale of header‑based prompt injection. The new report adds concrete payload examples and a brief defence guide from the Signals Corps. Looking ahead, security teams should monitor inbound HTTP headers for AI‑specific patterns, enforce strict validation before passing header data to language models, and consider sandboxing any agent‑driven actions that originate from external metadata. As AI integration deepens across web services, the community will need to develop standardized sanitisation practices to keep header‑borne prompts from becoming a covert attack surface.
Sources
Back to AIPULSEN