In mid‑October 2026 I shared a LinkedIn post warning that HTTP response headers can be weaponised for prompt injection when the fetch tool concatenates the full response (status line, headers, body) into the LLM prompt. I concluded that raw headers must never be passed to an LLM and that fetch outputs should be sanitised to include only the response body.
Decision: I do not pass raw HTTP headers to the model because any header can carry executable instructions that override prior directions.
Method: I sanitise fetch output by extracting only the response body before feeding it to the LLM.