A hash chain proves the ordering. Four attacks prove it's not enough

Description and use

Hash chains guarantee ordering but not authenticity. Four AI agent attacks were identified and fixed in NoireBox, validating the system’s reliability.

Problem

AI agents can self‑review, rewrite, farm or forge verdicts, compromising authenticity of audit logs.

Outcome

Three attacks were caught, one was provably detected; a deployment‑contract gap was closed during the post.

Workflow

• Clone NoireBox repo
• Install dependencies
• Run pytest on tests/test_scar_replay.py to replay four attacks
• Run demo_mcp_toolset.py to demonstrate lockfile

Evidence

A hash chain seals all four faithfully. That is the whole point: it proves the sequence happened in that order.

Source

SCAR-PROC-43, the rubber stamp → caught twice.

Source

SCAR-PROC-89, the farmed oracle → the omission, or the trail.

Source

SCAR-PROC-110, the forged verdict → provable, and a gap that closed in public.

Source