Back to skill

Security audit

技能盘点+健康体检 · Skill Inventory & Health Check

Security checks for vulnerabilities and agentic risk

Overview

This skill is a local skill-inventory tool with disclosed, user-triggered writes for disabling skills and saving a language preference; I found no hidden network use, autonomous persistence, exfiltration, or destructive behavior.

Install only if you are comfortable with a local tool that can read skill folders and related automation/config files. Treat any close candidates as advice, run impact review before disabling anything, and use --apply --yes only when you intentionally want it to persist changes to WorkBuddy skillOverrides.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (17)

Tp4

High
Category
MCP Tool Poisoning
Confidence
99% confidence
Finding

The declared description is for a skill inventory/health-check analyzer with mostly read-only behavior and only two specifically disclosed write paths. The supplied code instead implements a packaging tool for distribution: it collects whitelisted repository files, writes a zip archive under dist/, inspects the zip for excluded files and risky keywords, and validates packaging constraints. This is a materially different primary purpose and introduces routine file-writing behavior not covered by the declared write paths. Although it does not itself perform network or subprocess actions, it is clearly build/deployment-related rather than directory inventory and classification logic. Therefore the description does not accurately represent this code chunk.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 12)May include surrounding context.

md
allowed-tools: Bash(python scripts/skill_inventory.py:*), Read, Glob, Grep, Write(~/.workbuddy/settings.json), Edit(~/.workbuddy/settings.json), Write(~/.workbu

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 82)May include surrounding context.

md
allowed-tools: Bash(python scripts/skill_inventory.py:*), Read, Glob, Grep, Write(~/.workbuddy/settings.json), Edit(~/.workbuddy/settings.json), Write(~/.workbu

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 119)May include surrounding context.

md
allowed-tools: Bash(python scripts/skill_inventory.py:*), Read, Glob, Grep, Write(~/.workbuddy/settings.json), Edit(~/.workbuddy/settings.json), Write(~/.workbu

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 138)May include surrounding context.

md
allowed-tools: Bash(python scripts/skill_inventory.py:*), Read, Glob, Grep, Write(~/.workbuddy/settings.json), Edit(~/.workbuddy/settings.json), Write(~/.workbu

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 90)May include surrounding context.

md
| 技能目录 | ✅(已知平台可自动探测) | 每个子目录 = 一个技能(含 `SKILL.md` 即可);其他平台用 `--root <目录>`,或直接 `--probe` 让它自己找 |

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 176)May include surrounding context.

md
| 技能目录 | ✅(已知平台可自动探测) | 每个子目录 = 一个技能(含 `SKILL.md` 即可);其他平台用 `--root <目录>`,或直接 `--probe` 让它自己找 |

Lp1

High
Category
MCP Least Privilege
Confidence
75% confidence
Finding

The skill uses 'env' capability that is not listed in its permissions. This may indicate deceptive intent or missing permission declarations.

Content

No source excerpt is available for this finding.

Lp1

High
Category
MCP Least Privilege
Confidence
75% confidence
Finding

The skill uses 'file_write' capability that is not listed in its permissions. This may indicate deceptive intent or missing permission declarations.

Content

No source excerpt is available for this finding.

Lp1

High
Category
MCP Least Privilege
Confidence
75% confidence
Finding

The skill uses 'network' capability that is not listed in its permissions. This may indicate deceptive intent or missing permission declarations.

Content

No source excerpt is available for this finding.

Lp1

High
Category
MCP Least Privilege
Confidence
75% confidence
Finding

The skill uses 'shell' capability that is not listed in its permissions. This may indicate deceptive intent or missing permission declarations.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
80% confidence
Finding

The skill documents a persistent configuration write to ~/.workbuddy/settings.json that can disable skills by merging entries into skillOverrides, plus creation of a backup artifact. Even though this is consent-gated with --overrides --apply --yes, it is still a persistence-capable state change that survives the session and can alter future agent behavior if invoked mistakenly or through unsafe orchestration.

Content

Scanner excerpt · SKILL.en.md (reported line 212)May include surrounding context.

md
before trusting the tool. This section keeps only the **three SKILL-side technical points**
relevant to closing (no duplication of the README contract):

- **The tool never closes a skill on its own.** The close write path runs only with your
  explicit `--overrides --apply --yes` on a closable platform: it backs up `settings.json` first,
  then merges "off" entries into `skillOverrides`; without `--yes` it is a dry-run; non-closable
  platforms reject `--apply` outright.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.en.md (reported line 251)May include surrounding context.

md
| # | Wrong move | Why it's wrong | Right move |
|---|---|---|---|
| 1 | **Deleting the skill directory directly** to "close" a skill | Bypasses the backup artifact and protected-lock detection; may delete a skill still referenced by automations/hooks/triggers with no rollback | Let the tool emit an `--overrides` draft, then close via the host `/skills` menu or manual removal using the backup path printed in the report |
| 2 | **Closing without checking impact preview** | The skill may still be called implicitly by keyword triggers / experts / automations, leaving no log trace → false kill | Run `--impact <name>` first; confirm `referenced_by` is empty before acting |
| 3 | **Treating "no usage record" as "unused"** | "No record" only means no T1 explicit hit; T2 auto-mount / T3 keyword routing / T4 scheduled tasks / T5 hooks / T6 expert-internal calls leave no trace | Read both the "close candidates" and "manual review" buckets; confirm every candidate with the user before closing |
| 4 | **Expecting `--apply` to work on non-closable platforms** | `can_close=False` platforms (QwenWork / Baidu / generic) have no programmatic close channel; `--apply` is rejected outright | Use the host's own enable/disable toggles or manual client removal; the tool only reports inventory and candidates |
| 5 | **Treating `--protect` as a universal guard** | `--protect` only adds named skills to the protected bucket; it does not change the fact that "reverse-dependency scan only covers discoverable config roots" | Pair with `--refs` to point at extra reference-definition roots (automations/hooks/routing) for double coverage |

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

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.

Content

Scanner excerpt · SKILL.en.md (reported line 258)May include surrounding context.

md
| 6 | **`--root` pointing at a parent directory** that swallows unrelated dirs | In generic mode, `--root` scans **every** subdirectory under it as a skill, pulling in unrelated folders | Point `--root` exactly at the "skills root" (each subdirectory = one skill); on unknown hosts prefer `--probe` to let the tool find the root itself |

> Design posture: **under-report rather than false-kill.** The tool will never decide for you
> that a skill "may be closed" — it only emits candidates and impacts, and every write
> requires your explicit authorization plus a printed backup path.

---

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

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.

Content

Scanner excerpt · scripts/skill_inventory.py (reported line 17)May include surrounding context.

python
权限与驻留声明 / Permission & residency declaration
------------------------------------------------------
WRITE SCOPE: two target paths + one disclosed backup artifact, all declared here:
  1. ~/.workbuddy/settings.json -> key "skillOverrides", via --overrides --apply --yes
     (dry-run without --yes; see the backup artifact below). Written atomically:
     a temp file in the same directory, then os.replace -- a half-written config is

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

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.

Content

Scanner excerpt · scripts/skill_inventory.py (reported line 32)May include surrounding context.

python
its own config open) are retried twice at 0.5s before the error is surfaced.
ERROR CODES: E-LOG (usage log missing/unreadable) · E-ROOT (skills dir not found) ·
E-PLATFORM (unknown --agent) · E-NOCLOSE (host has no programmatic close channel) ·
E-CONF (settings.json read/write failed; nothing modified) · E-READ (one skill dir unreadable, skipped).
AUTONOMOUS RESIDENCY: none (no cron, startup script or daemon; no self-modification). This tool
performs no autonomous execution. It does perform two user-consented, reversible cross-session
writes — the "off" entries in skillOverrides (reversible via the host's /skills menu) and the

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The auto language resolution falls back to Chinese unconditionally when no explicit or saved preference is present. This imposes a specific locale by default rather than offering a neutral choice or explicit user opt-in, which conflicts with the language/locale policy criteria.

Content

No source excerpt is available for this finding.