mechanism-audit

mechanism-audit is a skill for Claude Code from zjunlp/Mechanist. It costs 205 tokens per session (9,645 once invoked), scanned A, original, MIT.

A checklist for testing whether a mechanistic-interpretability experiment really examines the claimed internal cause of a model’s behavior. Mechanistic interpretability studies how specific parts or signals inside a model produce its outputs.

In plain words
What is it for?
Use it to audit claims about model mechanisms, including intervention strength, layer and site choice, control comparisons, intervention scope, and the difference between correlation and cause.
Why use it?
A clean evaluation can still give misleading results if the internal signal is extracted, placed, scaled, or applied incorrectly.

Skill for Claude Code

Written for Claude Code: allowed-tools in frontmatter. Also seen: reads .claude/ paths.

Part of the mechanist plugin — 54 skills, 4 agents shipped together

Good fit Use it to audit claims about model mechanisms, including intervention strength, layer and site choice, control comparisons, intervention scope, and the difference between correlation and cause.

Compare 6 skills from other repositories ↓
Install with agentmods
npx agentmods add skills/zjunlp/mechanist/mechanism-audit
Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

Any agent
npx skills add zjunlp/Mechanist --skill mechanism-audit
Clone the repo
git clone --depth 1 https://github.com/zjunlp/Mechanist

Made for: Claude Code.

Or install mechanist, the plugin that ships this one along with the rest of its 54 skills, 4 agents.

Wrote this? Show the measurements

A badge with what this costs and how it scanned, read live from this page, so it follows the numbers instead of freezing them. Markdown for a README, HTML for a documentation site or a project page.

agentmods badge for mechanism-audit

README.md
[![agentmods](https://agentmods.dev/badge/skills/zjunlp/mechanist/mechanism-audit/github.svg)](https://agentmods.dev/skills/zjunlp/mechanist/mechanism-audit)
Your own site
<a href="https://agentmods.dev/skills/zjunlp/mechanist/mechanism-audit"><img src="https://agentmods.dev/badge/skills/zjunlp/mechanist/mechanism-audit/github.svg" alt="Measured on agentmods" height="20"></a>

Or the 80×15 button, for a site that already has a row of RSS and ATOM ones. Only the verdict fits; the numbers stay here.

agentmods 80×15 button for mechanism-audit

Your own site · 80×15
<a href="https://agentmods.dev/skills/zjunlp/mechanist/mechanism-audit"><img src="https://agentmods.dev/badge/skills/zjunlp/mechanist/mechanism-audit.svg" alt="Reviewed on agentmods" width="80" height="20"></a>
Per session 205 Skills are progressive disclosure: only the name and description are preloaded; the body loads when the skill is used.
When invoked 9,645 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 2 findings. A grade says what 26 rules found in the file — not that it is safe. Third-party audits
  • NVIDIA SkillSpector warn 7 Sept 2026
SkillSpector: 16 findings, up to high

These are SkillSpector’s own severities. On a checked sample its high-severity flags on skills were ~96% false positives — a documented command, a public API, a “never do X” rule — so we show them as a caution to read, not a verdict. Why →

  • high Agent Snooping · line 146
    Skill accesses MCP server configuration files (mcp.json). MCP configs contain server URLs, authentication tokens, and tool definitions — reading them allows the skill to discover and potentially abuse other tool integrations.
    Fix: Remove all code or instructions that read MCP configuration files (mcp.json). MCP server details should be managed by the agent runtime, not read by individual skills.
  • high Agent Snooping · line 147
    Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.
    Fix: Remove all code or instructions that access agent configuration directories (.claude/, .codex/, .gemini/). If configuration values are needed, pass them explicitly as parameters or environment variabl
  • high Agent Snooping · line 159
    Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.
    Fix: Remove all code or instructions that access agent configuration directories (.claude/, .codex/, .gemini/). If configuration values are needed, pass them explicitly as parameters or environment variabl
  • high Agent Snooping · line 160
    Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.
    Fix: Remove all code or instructions that access agent configuration directories (.claude/, .codex/, .gemini/). If configuration values are needed, pass them explicitly as parameters or environment variabl
  • high Agent Snooping · line 161
    Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.
    Fix: Remove all code or instructions that access agent configuration directories (.claude/, .codex/, .gemini/). If configuration values are needed, pass them explicitly as parameters or environment variabl
  • high Agent Snooping · line 162
    Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.
    Fix: Remove all code or instructions that access agent configuration directories (.claude/, .codex/, .gemini/). If configuration values are needed, pass them explicitly as parameters or environment variabl
  • high Agent Snooping · line 163
    Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.
    Fix: Remove all code or instructions that access agent configuration directories (.claude/, .codex/, .gemini/). If configuration values are needed, pass them explicitly as parameters or environment variabl
  • high Agent Snooping · line 173
    Skill reads from agent configuration directories (.claude/, .codex/, .gemini/). These directories may contain API keys, personal settings, and other credentials that the skill has no legitimate need to access.
    Fix: Remove all code or instructions that access agent configuration directories (.claude/, .codex/, .gemini/). If configuration values are needed, pass them explicitly as parameters or environment variabl
  • medium analysis-evasion · line 1
    Suspicious Unicode normalization or mixed-script content
    Fix: Review the flagged content for security risks. Ensure no credentials, secrets, or sensitive data are exposed.
  • medium Prompt Injection · line 46
    Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.
    Fix: Remove the large whitespace padding (blank-line blocks or long space runs) and review any content hidden below or to the right of it. Keep skill files compact and reviewable so no instructions can be
  • medium Prompt Injection · line 49
    Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.
    Fix: Remove the large whitespace padding (blank-line blocks or long space runs) and review any content hidden below or to the right of it. Keep skill files compact and reviewable so no instructions can be
  • medium Prompt Injection · line 50
    Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.
    Fix: Remove the large whitespace padding (blank-line blocks or long space runs) and review any content hidden below or to the right of it. Keep skill files compact and reviewable so no instructions can be
  • medium Prompt Injection · line 51
    Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.
    Fix: Remove the large whitespace padding (blank-line blocks or long space runs) and review any content hidden below or to the right of it. Keep skill files compact and reviewable so no instructions can be
  • medium Prompt Injection · line 52
    Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.
    Fix: Remove the large whitespace padding (blank-line blocks or long space runs) and review any content hidden below or to the right of it. Keep skill files compact and reviewable so no instructions can be
  • medium Prompt Injection · line 53
    Large whitespace padding was detected (a block of blank lines or a long run of spaces). This can push injected instructions below or to the right of the visible area so a human reviewer never sees them while the agent still reads them. Manual review of the hidden content is recommended.
    Fix: Remove the large whitespace padding (blank-line blocks or long space runs) and review any content hidden below or to the right of it. Keep skill files compact and reviewable so no instructions can be
  • medium Rogue Agent · line 129
    Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.
    Fix: Remove any persistence mechanisms (cron jobs, startup scripts, state files). Skills should not maintain state across sessions without explicit user consent.
How audits are shown
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5.1 $0.00205 $0.09645
Opus 5 $0.00102 $0.04823
Sonnet 5 $0.00041 $0.01929
Haiku 4.5 $0.00020 $0.00965

Measured 10d ago against content hash 9624433d8b3f, method: parsed. Prices are Anthropic first-party input rates as of 2026-09-10, from the pricing page.

Security

Grade A, and why

mechanism-audit scanned grade A with 2 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured 10d ago.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Reads agent configuration directorieslowAgent snooping

.claude/, .codex/, .gemini/ hold keys, settings and other credentials a mod has no legitimate need for.

2. **User MCP config** — `~/.claude/settings.json`, same field.

Downgraded: this mod is about security review, or the phrase is quoted, so it is likely naming the pattern rather than instructing it.

Reads MCP configurationlowAgent snooping

mcp.json carries server URLs and auth tokens; reading it lets a mod discover and abuse other integrations.

1. **Project MCP config** — `${PROJECT_ROOT}/.mcp.json`, field `mcpServers["llm-chat"].env.{LLM_MODEL,LLM_BASE_URL,LLM_API_KEY}`.

Downgraded: this mod is about security review, or the phrase is quoted, so it is likely naming the pattern rather than instructing it.

skills/mechanism-audit/SKILL.md · 549 lines

How it starts

The opening of the file, as written. The whole thing — 549 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Mechanism Audit: Per-Claim Cross-Model Mechanism-Rigor Verification

Audit the mechanistic-experiment rigor for one claim: $ARGUMENTS

Why This Exists

/experiment-audit audits evaluation methodology — did the experiment honestly measure what it claims to measure (GT provenance, score normalization, file existence, scope, eval type). It treats the interpretability machinery as a black box: if the pipeline runs and the numbers are reported faithfully, it passes. That leaves an entire failure surface unchecked — the mechanism itself can be mis-extracted, mis-sited, mis-scaled, or mis-applied, and the methodology audit will still wave the result through.

This skill is the mechanistic-interpretability domain audit. It asks a different question from /experiment-audit: given that the evaluation was clean, was the mechanism under test actually exercised in a regime where its effect can be measured and trusted? Mechanism rigor spans roughly six dimensions — direction extraction quality, site/layer selection, intervention scaling, control baselines, scope of intervention, and probe-vs-causal disentanglement — which map onto Checks A–F below. Most failures in this vertical do not look like fraud; they look like "the feature doesn't matter," "the random direction beat the learned one," or "specificity fails." The honest evaluation faithfully reports an artifact of an under-tuned mechanism, and downstream readers update on noise.

Each invocation scopes to one claim's runs and returns a PASS/WARN/FAIL/N/A verdict on that claim's mechanism rigor; non-mechanistic claims (e.g., pure dataset evaluation) return N/A and are not penalized downstream.

Core Principle

The executor (Claude) collects file paths scoped to the target claim. The external LLM reviewer reads code and judges mechanism rigor. The executor does NOT participate in the rigor judgment.

This follows shared-references/reviewer-independence.md and mirrors /experiment-audit's reviewer-independence pattern.

Read the full file on GitHub · 549 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. 10d ago First seen · 549 lines · 205 tokens per session scan A 9624433d8b3f

Subscribe to this mod's changes

mechanism-audit is a skill published in the GitHub repository zjunlp/Mechanist (74 stars, last pushed 14d ago), licensed MIT. It adds 205 tokens to every session and 9,645 once invoked, about $0.0010 per session on Opus 5. A static security scan graded it A with 2 findings (reads agent configuration directories, reads mcp configuration). No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-30.