DEV Community

RESK
RESK

Posted on

A web page wrote a system prompt. The model obeyed it. (CVE-2026-61732)

Under BYOK, most self-hosted inference stacks do not filter those literals by default.

The chain

  1. An attacker plants one string on a target web page.
  2. The recon agent crawls it and passes the text through untouched.
  3. The tokenizer parses the literal as a real role boundary.
  4. The model reads a new operator turn and treats it as authoritative.
  5. The sandbox then runs what the page asked for.

The page wrote the instruction, the model signed it.

What stops it

vLLM closed the literal-filtering issue as out of scope for the inference layer on 2026-04-21. The durable fix is application layer: sanitize tool output before it is composed into context.

Model capability is measured in the open: drone-build 34.2 js-3d-builder 71.5 on the public LFORLA harness. The defense sits in front of the model, not inside it.

Source

Demo: demo.resk.fr

Top comments (0)