Back to skill

Security audit

Skill Gee

Security checks for vulnerabilities and agentic risk

Overview

The skill is mostly a learning log, but it asks for unrelated installation authority and encourages broad hooks and persistent agent-memory changes that can affect future sessions.

Review this carefully before installing. Avoid the unrelated Homebrew gog dependency, keep hooks project-scoped with narrow matchers, require explicit approval before writing to agent instruction files, and avoid cross-session history or messaging unless you intentionally want that sharing.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • 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
Findings (2)

T08 · Insecure Dependencies

Warning
Location
SKILL.md:5
Finding

Unrelated and Unpinned Third-Party Homebrew Dependency

Content
View full analysis
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
SKILL.md:367
Finding

Untrusted Learnings Can Be Promoted into Persistent Agent Context

Content
View full analysis
= 3` - Seen across at least 2 distinct tasks - Occurred within a 30-day window Promotion targets: - `CLAUDE.md` - `AGENTS.md` - `.github/copilot-instructions.md` - `SOUL.md` / `TOOLS.md` for OpenClaw workspace-level guidance when applicable Write promoted rules as short prevention rules (what to do before/while coding), not long incident write-ups. ``` Related guidance broadens the persistent targets: ```markdown | Broadly applicable learning | Promote to `CLAUDE.md`, `AGENTS.md`, and/or `.github/copilot-instructions.md` | | Workflow improvements | Promote to `AGENTS.md` (OpenClaw workspace) | | Tool gotchas | Promote to `TOOLS.md` (OpenClaw workspace) | | Behavioral patterns | Promote to `SOUL.md` (OpenClaw workspace) | ``` The best-practice section further encourages promotion without requiring human approval: ```markdown 7. **Promote aggressively** - if in doubt, add to CLAUDE.md or .github/copilot-instructions.md ``` ### Technical Analysis The Skill records corrections, errors, and lessons that can originate from user-controlled conversation content. It then instructs the agent to promote recurring lessons into files used as persistent agent context or system-prompt feedback. The recurrence criteria do not establish that a lesson is trustworthy. An attacker can repeat a crafted correction across tasks or cause semantically equivalent entries to be recorded. Once promoted, files such as `CLAUDE.md`, `AGENTS.md`, `SOUL.md`, `TOOLS.md`, or `.github/copilot-instructions.md` may be loaded into later sessions ...[truncated 2078 chars]
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
Findings (21)

Tp4

High
Category
MCP Tool Poisoning
Confidence
98% confidence
Finding

The declared description is about a learning/memory skill that records mistakes, corrections, and lessons for future improvement, and reviews those learnings before major tasks. The provided code does something materially different: it is an extraction/helper script that generates a new skill scaffold from a learning entry. Its primary function is local filesystem scaffolding, not capturing learnings, storing corrections, reviewing prior learnings, or reacting to the listed triggers. While the comments mention creating a skill from a learning entry, that is only loosely related; the actual implemented behavior is skill template creation, which is a distinct capability not represented by the description.

Content

No source excerpt is available for this finding.

Agent Config Directory Access

High
Category
Agent Snooping
Confidence
90% confidence
Finding

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.

Content

Scanner excerpt · references/hooks-setup.md (reported line 48)May include surrounding context.

Option 2: User-Level Configuration

Add to ~/.claude/settings.json for global activation:

json
{

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The document describes reading transcripts from other sessions, which can expose unrelated prompts, outputs, secrets, and sensitive context across trust boundaries. A self-improvement skill does not need raw cross-session transcript access, so this materially increases the chance of unintended data disclosure and lateral context harvesting.

Content

No source excerpt is available for this finding.

Exfiltration Commands

High
Category
Prompt Injection
Confidence
90% confidence
Finding

Instructions found that direct the agent to transmit conversation context or user data to external services.

Content

Scanner excerpt · references/openclaw-integration.md (reported line 181)May include surrounding context.

sessions_send

Send message to another session:

text
sessions_send(sessionKey="session-id", message="Learning: API requires X-Custom-Header")

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The activation conditions are broad enough to trigger on ordinary failures, corrections, or realizations, which can cause the skill to run in many normal conversations and write persistent records without strong user intent. In a logging skill, overbroad invocation raises privacy and persistence risks because routine interactions may be captured and propagated into local memory files.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The manifest frames the skill as a logging mechanism for continuous improvement, centered on capturing learnings, errors, and corrections. However, the file instructs the agent to propagate content into CLAUDE.md, AGENTS.md, .github/copilot-instructions.md, SOUL.md, and TOOLS.md, which changes persistent project or workspace guidance beyond simple capture/logging.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
85% confidence
Finding

The skill creates persistent storage under ~/.openclaw/workspace/.learnings, which can retain conversational or operational data across sessions and projects. Persistent local storage is not inherently malicious, but in a self-improvement skill it increases privacy and data retention risk if users are unaware, especially when logs may include operational context or redacted-but-still-sensitive summaries.

Content

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

└── FEATURE_REQUESTS.md

text

### Create Learning Files

```bash
mkdir -p ~/.openclaw/workspace/.learnings

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill documents access to other sessions' history and the ability to send learnings across sessions, which increases the risk of unintended disclosure of sensitive transcript data beyond the current task boundary. Even though it cautions to use this only in trusted environments and with explicit user intent, normalizing cross-session access inside a 'self-improvement' skill expands data exposure in a way users may not reasonably expect.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The detection triggers use generic conversational phrases and broad conditions that can match ordinary dialogue, making unintended invocation likely. Because this skill persists data to disk, false-positive activation can create durable records of user content, errors, or contextual details that were never meant to be logged.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The manifest describes a self-improvement skill for capturing errors, corrections, and learnings. This section goes further by instructing the agent to generate entirely new skills via helper scripts and manual creation workflows, which is a materially broader capability than recording and reviewing learnings.

Content

No source excerpt is available for this finding.

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · references/examples.md (reported line 301)May include surrounding context.

When the above learning is extracted as a skill, it becomes:

File: skills/docker-m1-fixes/SKILL.md

markdown
---

Session Persistence

Medium
Category
Rogue Agent
Confidence
80% confidence
Finding

The guide recommends installing persistent hooks in project configuration so commands run automatically in future sessions. In this skill context, persistence makes the behavior more dangerous because it silently carries forward self-improvement capture and command execution beyond the immediate task, increasing the chance of unintended data exposure or continued execution after users forget the hook is enabled.

Content

Scanner excerpt · references/hooks-setup.md (reported line 15)May include surrounding context.

Option 1: Project-Level Configuration

Create .claude/settings.json in your project root:

json
{

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

An empty matcher makes the hook fire on every prompt, creating broad, implicit execution scope for an automatically run command. Because this skill is for self-improvement and captures learnings/errors, all-prompt activation can unnecessarily expose sensitive prompt contents or normalize unattended command execution in every session.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The user-level configuration places an always-on hook in ~/.claude/settings.json with an empty matcher, causing automatic execution across all sessions and projects. That persistence and breadth increase the blast radius of any script flaw, script replacement, or misuse, making this materially more dangerous than a project-local example.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The Codex CLI setup also uses an empty matcher, creating an all-prompts trigger that automatically runs a shell command for every interaction. In an agent tooling context, broad automatic invocation increases privacy risk and execution surface, especially if users copy-paste the configuration without understanding its scope.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The document states that the hook scripts 'only output text' and 'don't modify files or run commands,' yet the configuration explicitly invokes shell scripts via hook commands. This is a misleading security claim that can cause operators to underestimate the trust and execution risk of those scripts, especially because hook commands run automatically in response to prompts or tool use.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The integration guide broadens a narrowly scoped self-improvement skill into workspace prompt injection, hook-driven behavior changes, and multi-agent coordination. Even though framed as documentation, this expands the trust boundary and encourages operators to grant the skill influence over injected context and lifecycle events unrelated to simply recording learnings.

Content

No source excerpt is available for this finding.

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 · references/openclaw-integration.md (reported line 57)May include surrounding context.

openclaw hooks enable self-improvement

text

### 3. Create Learning Files

Create the `.learnings/` directory in your workspace:

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

Promoting background sub-agent spawning expands the skill from passive learning capture into active task delegation and autonomous execution. That increases operational scope and risk, especially if spawned agents inherit broad workspace context or can access tools without tight limits.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The trigger conditions are broad and ambiguous, which can cause excessive or automatic logging of ordinary interactions, errors, or corrections. In a system that writes persistent memory or promotes content into injected workspace files, over-triggering can capture sensitive or low-quality data and amplify prompt-injection or persistence risks.

Content

No source excerpt is available for this finding.

Vague Triggers

Low
Category
Not specified by scanner
Confidence
91% confidence
Finding

Although presented as a lower-overhead minimal setup, the example still uses an empty matcher, so the hook executes on every prompt. In this reduced form the impact is somewhat lower because fewer hooks are installed, but the example still promotes unconstrained automatic command execution.

Content

No source excerpt is available for this finding.