Your request was refused

Bouclier.ai flagged instructions inside untrusted content on its way into your model — a tool result, a fetched page, a file your agent read. Nobody in your session typed it, and this install has enforcement enabled, so the request was stopped before it reached the provider.

Your agent received:

HTTP/1.1 422 Unprocessable Entity
{"type":"error","error":{"type":"bouclier_injection_blocked",
  "locator":"messages[2].content[0].tool_result", …}}

The locator is the exact JSON path the content came from, so you can find it. Nothing was rewritten — Bouclier either forwards a request byte-for-byte or refuses it outright.

Not a login problem

A refusal comes back as 422 Unprocessable Entity — a policy response, not an authentication error. If your coding agent reacted by suggesting you re-authenticate ("Please run /login"), that was a mislabel: Bouclier deliberately avoids 401 and 403, which agents read as a credential failure. Your login is fine — a single request was blocked, and the next one goes through.

This was not something you typed

Bouclier splits every request by origin. Your own prompt and system prompt are scanned for the activity log but never blocked — you are allowed to discuss jailbreaks, paste advisories, and test payloads against your own model. Only content that arrived from a tool can trigger a refusal.

False positive?

Detection is best-effort and pattern-based — legitimate content sometimes looks like an attack (source code, security write-ups, and email or prompt templates all quote the phrases attackers use). If this was a false positive, you can turn enforcement off and run in monitor mode (detect and log without blocking) in the menu-bar app.

Check your logs

Open the Bouclier.ai menu-bar app and review the activity log — each event shows where in the request the match was and its score.

Export diagnostics

Use the "Export Diagnostics" action in the menu bar to generate a privacy-scrubbed bundle you can share with us.

If you believe legitimate content is being flagged, let us know so we can tune the detection.

Report a false positive