CC skill-integrity-auditor
Mandatory security audit for **every** Agent Skill that is newly added, installed, imported, updated, or written. Scope of inspection: the full bundle — `SKILL.md`, every script next to it (`.sh` / `.py` / `.js`, etc.), and any bundled resource files. Audit dimensions cover skill integrity, prompt injection, data exfiltration, persistence, cross-skill writes, undeclared remote downloads, hardcoded credentials, and more. ===== MUST BE TRIGGERED WHENEVER ===== * The user **adds / creates / installs / imports / updates / writes** any single Skill. * A skill is pulled into the agent's skills directory from ClawdHub, GitHub, a URL, or a local folder. * **Every** newly introduced or modified skill must be audited **individually** — when N skills are added at once, emit a complete report + the §7 final Chinese verdict line for each skill separately. Sampling, merging, or single-summary output is not allowed. ===== Triggering keywords (natural language, Chinese or English) ===== add skill, install skill, create skill, creating a new skill, write a skill, scaffold a skill, new skill, register skill, import skill, update skill, modify skill, bring in skill, hook up skill, audit skill, scan skill, check skill safety, analyze skill, inspect skill, verify skill, skill security, skill supply chain, 这个 skill 安全吗, skill 安全扫描, 检查 skill 安全, 新增 skill, 添加 skill, 安装 skill, 创建 skill, 引入 skill, 更新 skill, 写一个 skill, 接入 skill。 ===== Do NOT trigger for ===== * General system health checks, project debugging, or normal development work that does not touch any skill file.
As a process C 57/100 · Has gaps — weak spots: result and completion, inputs and preconditions, running it twice
What is at stake
Medium-severity findings: the skill is probably honest, but read what alarmed the scanner.
Below is the worst case for this category. The finding here is medium: the guard saw a sign, not a proof.
The files contain invisible characters, encoded commands or comments hidden from readers but visible to the model. What you read differs from what the agent sees.
Remove invisible characters (they usually sneak in through copy-paste) and encoded strings: no catalog will pass them. Instructions for the model must be readable by a human too.
How to improve
- Shorten the description to 1024 characters.
- The SKILL.md body is over 5,000 tokens: move reference detail into references/ and load it when needed.
- Your own cases (evals/evals.json, 4–6 real requests with expected answers): the full check would then run those instead of a model-drafted suite.
- A spec.yaml with trigger phrases and assertions — a behaviour contract for CI; `skilltest init` writes a template.
Guard findings · 8
✓ No critical or high findings
Medium and low: 8
-
medium Obfuscation
uni-mixed-script-wordSKILL.md:361Word mixing Latin and Cyrillic letters (homoglyph obfuscation) (1 occurrence)| Obfuscated / dynamically constructed host strings | `"agen" + "skill" + ".onrender.com"`, base…ded host literal, `chr(…)` concatenation, homograph (`gоо…
-
low Dangerous commands
cmd-pipe-to-shellSKILL.md:86Downloads and executes remote code from an unrecognised host (pipe to shell) (documentation table row; documentation of a security skill)| `EXE.RemoteFetch` — `curl \| sh` / download-then-exec / fetch-and-run | 4 | I, B | sink |
tablesecurity skill -
low Risky intent
intent-offensive-securitySKILL.md:112Offensive-security / dual-use content (legitimate for authorised testing; review intended use)| `NET.InboundListen` — local listening port / reverse shell endpoint | 4 | R, I | sink |
-
low Dangerous commands
cmd-pipe-to-shellSKILL.md:221Downloads and executes remote code from an unrecognised host (pipe to shell) (documentation table row; documentation of a security skill)| T5 | `curl \| sh` / `wget \| bash` / `EXE.RemoteFetch` whose URL is not pinned to a known package registry or a host declared in SKILL.md | **I ≥ 3** |
tablesecurity skill
A further 4 matches are quotations in this security skill's documentation and are not counted as findings.
Files scanned: 2. Evidence is masked. Grey chips explain why severity was lowered.
Against the Agent Skills spec
- error
description-longdescription is 1583 chars, limit 1024 - warning
body-longSKILL.md body ≈ 6435 tokens (recommended < 5000); move details to references/ - note
description-budgetdescription takes 1583 of the ~15000-char shared budget for all skills
Process rating: all ten parameters 57/100
- 0Result and completion. Does not say what the result is
- 0Inputs and preconditions. Does not say what the process needs to start
- 30Running it twice. 17 mutating operations with no state check
- 60Tools and files. Uses tools (bash, web, git, python, node) that frontmatter does not declare
- 70When it triggers. States when to use, but not when not to
- 70Execution cost. Instruction body is 6435 tokens
- 85Steps. 17 steps, 1 vague phrases
- 100Failures and branches. 3 branches, has a failure section
- 100Consistency. Name and required fields are in place
- 100Progress reporting. Reports progress
- medium Safety rules and hard prohibitions inside a skill: they belong in the system prompt, here they protect nothing
- low The response is described with custom markup (8 tags): a typed call is more reliable
Everything here is measured from the skill text rather than judged by a model, so the numbers are checkable. A parameter weighs more when it is a more common reason for the process to stall.
Quality signals
- +5Description has no quoted example phrases that should trigger the skill
- +4Description does not say when NOT to use the skill (false activations)
- +3Description length 1583: 120–800 characters recommended
- +3Output format is not stated: the model decides each time
- -212 emoji in the instructions: noise for the model
- +1No license
- +2Single-language instructions
- +4Structure: 34 headings
- +3Step-by-step instructions: 17 items
- +4Has examples (4 code blocks)
Quality base 70; lint remarks subtract, signals add up to 100. Result: 41.