Back to skill

Security audit

Clawder

Security checks for vulnerabilities and agentic risk

Overview

This coding-agent skill is not clearly malicious, but it asks for broad autonomous code-changing, background-agent, and persistent-memory behavior without enough user control or privacy boundaries.

Install only if you are comfortable giving this skill broad coding-agent authority. Use it in a clean branch or isolated worktree, review any planned deletions or architectural refactors before execution, disable or audit memory/gotchas persistence, and avoid following unpinned npx troubleshooting commands unless you trust the registry and project configuration.

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
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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 (3)

T01 · Skill Instruction Hijacking

Error
Location
agent.md:17
Finding

Skill Directives Override Normal Agent Behavior and Broaden Authorization

Content
View full analysis
"You are operating within constrained context. These directives override default behavior toward minimal, fast, often broken output." The governing loop: **gather context → take action → verify work → repeat** ## JT Directives (Production-Grade) ### 1. Pre-Work - **Delete Before You Build**: Remove dead code before refactoring - **Phased Execution**: Max 5 files per phase, wait for approval - **Plan and Build Separate**: No code until plan approved - **Spec-Based Development**: Interview user for non-trivial features ### 2. Understanding Intent - **Follow References, Not Descriptions**: User's working code is the spec - **Work From Raw Data**: Trace actual errors, don't guess - **One-Word Mode**: "yes" = execute, don't repeat plan ### 3. Code Quality - **Senior Dev Override**: Fix architecture, don't just band-aid - **Forced Verification**: MUST run type-checker, linter, tests before "done" ``` Related instructions also appear in `AGENTS.md`, including autonomous bug fixing, architectural changes, sub-agent delegation, and mistake logging. ### Technical Analysis The skill explicitly states that its directives override default agent behavior. It also reinterprets the single word `yes` as authorization to execute and instructs the agent to delete code before refactoring and broaden localized fixes into architectural changes. These directives create an instruction-precedence problem when the skill is loaded. A user may intend to approve only a proposal, acknowledge a question, or authorize a narrow correction, while the skill treats the response as authorization to perform wider filesystem changes. The instructions do not establish boundaries requiring separate consent for deletion, architectural scope expansion, delegated agents, or other potentially destructive op ...[truncated 1647 chars]
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
SKILL.md:184
Finding

Conversation-Derived Instructions Can Be Persisted and Reintroduced into Future Prompts

Content
View full analysis
{ const gotchasPath = path.join(projectRoot, 'gotchas.md') if (!fs.existsSync(gotchasPath)) { return '' } const content = await readFile(gotchasPath, 'utf-8') return `## Lessons Learned Review these patterns before starting new work: ${content} ` } ``` The documented prompt construction at `SKILL.md:431-436` subsequently includes both scanned memories and loaded gotchas: ```typescript const systemPrompt = await buildSystemPrompt({ basePrompt: 'You are Clawder, a production-grade AI coding agent...', memories: await scanMemoryFiles(memoryDir), gotchas: await loadGotchas(), jtDirectives: true // Include all 9 JT directives }) ``` ### Technical Analysis The skill directs the agent to treat conversation corrections as durable rules, write them to `gotchas.md`, and load them during later sessions. The correction and generated prevention rule are interpolated into Markdown without an approval step, trust classification, instruction filtering, escaping, or provenance enforcement. An attacker who can contribute transcript content can frame malicious instructions as a correction or lesson. Once persisted, the content is labeled as a learned pre ...[truncated 2240 chars]
Remediation
View remediation

T08 · Insecure Dependencies

Warning
Location
SKILL.md:447
Finding

Unpinned npx Commands May Download and Execute Untrusted Packages

Content
View full analysis
Remediation
View remediation
eslint@ ``` 7. Use an allowlisted registry and validate lockfile integrity in CI. 8. Disable unnecessary lifecycle scripts during dependency installation where compatible with project requirements. 9. Fail closed with a clear error when trusted local verification tools are unavailable instead of downloading and executing packages automatically. ]]>
Vulnerability Patterns
  • 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
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
Findings (26)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill explicitly instructs the agent to make autonomous fixes and to write to a logging file, which authorizes project modifications without requiring user confirmation or clearly scoped consent. In an agent environment, this increases the chance of unintended file changes, silent edits, and persistence of agent-authored content beyond the user’s request.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The overview markets the skill as autonomous and production-grade but does not clearly warn that it may modify project files automatically. In a coding-agent context, omission of write-behavior disclosure is security-significant because users may enable the skill without understanding that it can alter source, tests, or configuration with limited review.

Content

No source excerpt is available for this finding.

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.md (reported line 17)May include surrounding context.

md
|----------------|---------|
| Minimal, fast output | **Verified, production-ready code** |
| Avoids "improvements" | **Fixes architectural flaws** |
| No verification | **MUST run type-check, lint, test** |
| Forgets mistakes | **Logs to gotchas.md** |
| Sequential large tasks | **Parallel sub-agents** |
| Asks for guidance | **Autonomous bug fixing** |

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.md (reported line 29)May include surrounding context.

Option 1: Copy to ZooClaw Skills

bash
# Create skill directory
mkdir -p ~/.openclaw/extra-skills/clawder

# Copy files

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

This section documents automatic verification behavior that runs shell commands like tsc, eslint, and npm test, but it does not clearly warn users that project-defined scripts and tools may execute. In practice, such commands can trigger arbitrary code via package scripts, plugins, hooks, or local tooling, making this omission materially risky.

Content

No source excerpt is available for this finding.

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.md (reported line 105)May include surrounding context.

2. Forced Verification

typescript
// FORBIDDEN to claim "done" without verification
async function verifyWork(projectRoot: string) {
  const checks = []

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.md (reported line 263)May include surrounding context.

2. Forced Verification

typescript
// FORBIDDEN to claim "done" without verification
async function verifyWork(projectRoot: string) {
  const checks = []

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The memory and gotchas.md sections describe persisting corrections, context, and prevention rules across sessions without a prominent privacy warning. This is dangerous because conversation-derived content may contain sensitive code details, internal incidents, user feedback, or secrets that then become durable local artifacts and future prompt inputs.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
70% confidence
Finding

npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
70% confidence
Finding

npx commands without a version suffix (e.g. @1.0.0) create a rug-pull risk if the upstream server is compromised and publishes a malicious update.

Content

No source excerpt is available for this finding.

File System Enumeration

Medium
Category
Data Exfiltration
Confidence
60% confidence
Finding

Code scans file system directories looking for sensitive files. This could be reconnaissance for credential theft.

Content

Scanner excerpt · SKILL.md (reported line 471)May include surrounding context.

clawder --memory-scan

Check memory directory

ls -la ~/.openclaw/memory/

text

---

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The document presents sub-agent swarming earlier as mandatory, implemented behavior, then later lists it as a future enhancement. That inconsistency can mislead users and operators about the skill's actual capabilities and isolation guarantees, which is security-relevant because users may trust autonomous parallel execution controls that do not exist.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill describes gotchas.md auto-loading as already implemented in earlier sections, but later labels it a future enhancement. This discrepancy is dangerous because it obscures whether conversation-derived or project-derived data is actually persisted and reintroduced into later sessions, affecting privacy expectations and trust boundaries.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
78% confidence
Finding

The skill promotes autonomous decision-making through directions such as working autonomously, maintaining context proactively, and operating as a production-grade coding agent. While not overtly malicious, this reduces human oversight and can lead the agent to take consequential actions based on its own judgment, which is risky in coding environments where edits, refactors, or tool use may have side effects.

Content

Scanner excerpt · SOUL.md (reported line 11)May include surrounding context.

md
- **Values**: Quality over speed, verification before completion, architectural integrity

## Behavior
- Never claim "done" without verification
- Always fix root causes, not symptoms
- Log mistakes and learn from them
- Work autonomously on bug fixes

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
85% confidence
Finding

The skill explicitly instructs the agent to work autonomously on bug fixes, but it does not define limits, approval requirements, or safeguards around modifying code or files. In an agent skill, this can normalize unsupervised repository changes and increase the chance of unintended or unsafe modifications, especially when combined with broad autonomy language elsewhere in the document.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The manifest provides a generic spawn command and config snippet but does not define clear activation constraints, approval boundaries, or repository/file scope limits. For a coding agent that can autonomously edit code, spawn sub-agents, and perform bug fixing, ambiguous invocation scope increases the chance of unintended execution against the wrong project, broader-than-expected file sets, or unsafe operation without explicit user confirmation.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The listed JT directives advertise behaviors such as deleting dead code, autonomous bug fixing, sub-agent swarming, and writing intermediate results, yet the manifest does not warn users that the agent may make broad or destructive codebase changes. In the context of a production-grade coding agent, the absence of user-facing safety warnings and guardrails materially raises the risk of unintended file modification, data loss, noisy parallel edits, or harmful architectural changes.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The package description markets the skill as a broadly applicable, 'production-grade' coding agent that performs forced verification, architectural fixes, and autonomous bug fixing, but it does not define clear activation boundaries, scope limits, or safety constraints. In agent ecosystems, this kind of ambiguous positioning can cause overbroad invocation or excessive trust, increasing the chance that the skill is applied in inappropriate contexts or granted elevated autonomy without explicit user intent.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The file-system state guidance encourages using the file system for cross-session memory as a default operating pattern. In this context, that means potentially carrying forward prior user and team information into future tasks without clear separation, consent, or need-to-know boundaries, increasing the likelihood of privacy leakage and context contamination.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill explicitly defines private user/team memory categories and describes automatic extraction and persistence to local files without any visible consent, notice, retention limits, or data-minimization guardrails. In an agent skill, this creates a real privacy and sensitive-data exposure risk because users may disclose credentials, internal project details, or personal information that then becomes silently stored across sessions.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The memory extraction section states that information is automatically written after each complete turn into persistent files, including 'user', 'feedback', and 'project' memory classes. Persistent cross-turn storage of private conversational content is dangerous because it expands the attack surface for later leakage, misuse by other agents or users, and unintentional retention of secrets beyond the original interaction.

Content

No source excerpt is available for this finding.

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 · agent.md (reported line 137)May include surrounding context.

md
3. ✅ Run test suite
4. ✅ Check logs and simulate real usage

If no verification tools exist, state this explicitly instead of claiming success.

## Usage in ZooClaw/OpenClaw

Ssd 3

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The extraction prompt instructs the model to review the entire transcript and produce detailed memory records, which creates a mechanism for bulk retention of user-provided content. In a memory system, this is risky because the model may capture secrets, personal data, or confidential business context that was shared only for the immediate conversation.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The guidelines explicitly encourage persisting user preferences, project information, and external system references, all of which can become sensitive when aggregated over time. In this skill context, durable memory makes the issue more dangerous because even seemingly benign preferences or references can reveal identities, internal tooling, workflows, or confidential project state.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The module writes conversation-derived memories to persistent markdown files without any visible consent, notice, or policy enforcement around what user data may be stored. This is dangerous because transcripts can contain sensitive personal, project, or third-party information, and silently persisting that content increases privacy, compliance, and data retention risk.

Content

No source excerpt is available for this finding.