RedThread: An Open-Source CLI for Replayable LLM-Agent Red-Team Cases

Apprehensive-Zone148 · reddit · 2026-09-11

A developer has open-sourced RedThread, a small CLI focused on a narrow agent-security evaluation problem: untrusted text can alter a later tool call even when the final answer looks normal.

RedThread preserves the context slice, tool schema, proposed call, response, and score as a replayable case, so the same scenario can be rerun after changing the prompt, model, or adapter. The author stresses that a passing replay isn't proof of production safety and a denied call isn't proof of no impact — the goal is to make security claims testable rather than anecdotal.

Source: github.com/matheusht/redthread

Related event: RedThread: Open CLI Tool Makes Agent Red-Teaming Reproducible(2 posts)→

Original post →

More from coding & agent

coding & agent channel →