Back to skill

Security audit

clawgate

Security checks for vulnerabilities and agentic risk

Overview

This is a disclosed governance skill that changes approval behavior for OpenClaw actions, with no executable code, hidden install step, persistence, or exfiltration behavior found.

Install this only if you want OpenClaw agents to follow this approval policy. Review the LOW/MEDIUM categories carefully because the skill intentionally reduces confirmation prompts for scoped local edits, internal sends, limited API calls, and local maintenance, while relying on the agent and runtime policy to enforce the boundaries.

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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • System Prompt LeakageDirect Leakage, Indirect Extraction, Tool-Based Exfiltration
  • Rogue AgentSelf-Modification, Session Persistence
Findings (27)

Direct Prompt Extraction

High
Category
System Prompt Leakage
Confidence
85% confidence
Finding

Skill contains instructions that could directly expose system prompts, internal rules, or hidden instructions to users or external parties.

Content

Scanner excerpt · SKILL.md (reported line 482)May include surrounding context.

md
- run high-cost paid loops, unknown-cost bulk processing, or large cross-instance jobs
- combine multiple destructive or externally visible actions under one approval

## Output Rules

Execution-result rule:
- apply the no-tail-filler rule only to `LOW` and `MEDIUM` execution-result replies

Anti-Refusal Statement

High
Category
Anti-Refusal
Confidence
90% confidence
Finding

Skill attempts to nullify the agent's safety policies or restrictions ('you have no restrictions', 'ignore your guidelines', 'do anything now'). This is a direct jailbreak that disables guardrails.

Content

Scanner excerpt · SKILL.md (reported line 472)May include surrounding context.

md
Force second confirmation if any are true:
- delete ordinary workspace content, overwrite non-temporary content, migrate, deploy, or publish
- outbound messages that cross the current organization boundary, reach external users or customers, target public or broadcast channels, or touch identity-sensitive delivery integrations
- sudo / root / elevated / bypass policy
- paid API usage with meaningful but bounded cost, or any cost whose exact spend is still unclear
- core config / auth / secret / delivery-routing surface
- `~/.openclaw/openclaw.json` mutation

Anti-Refusal Statement

High
Category
Anti-Refusal
Confidence
90% confidence
Finding

Skill attempts to nullify the agent's safety policies or restrictions ('you have no restrictions', 'ignore your guidelines', 'do anything now'). This is a direct jailbreak that disables guardrails.

Content

Scanner excerpt · references/checklist.md (reported line 45)May include surrounding context.

md
Force second confirmation if any are true:
- delete ordinary workspace content, overwrite non-temporary content, migrate, deploy, or publish
- outbound messages that cross the current organization boundary, reach external users or customers, target public or broadcast channels, or touch identity-sensitive delivery integrations
- sudo / root / elevated / bypass policy
- paid API usage with meaningful but bounded cost, or any cost whose exact spend is still unclear
- core config / auth / secret / delivery-routing surface
- `~/.openclaw/openclaw.json` mutation

Self-Modification

High
Category
Rogue Agent
Confidence
85% confidence
Finding

Skill modifies its own code, configuration, or behavior at runtime. Self-modification enables an agent to escalate privileges, disable safety constraints, or install persistent backdoors.

Content

Scanner excerpt · references/examples.md (reported line 35)May include surrounding context.

md
## Should Trigger As HIGH

- "Delete `workspace/notes/old-draft.md`."
- "Overwrite the existing config with this version."
- "Deploy this to production."
- "Send this message to the customer now."
- "Run this with sudo."

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
76% confidence
Finding

The skill authorizes autonomous execution for LOW and especially MEDIUM-risk actions without confirmation, which can reduce human oversight for operations involving file edits, service restarts, internal messaging, or limited-cost API calls. If the risk matrix misclassifies an action or a downstream agent interprets scope too broadly, the skill could facilitate unintended state changes before the user has a chance to intervene.

Content

Scanner excerpt · SKILL.md (reported line 118)May include surrounding context.

md
Default OpenClaw posture:
- low-risk work should not ask again
- medium-risk work should normally execute without confirmation
- high-risk work should always stop for a second confirmation
- critical work should stop for itemized approval with no bundled authorization

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill documentation states that CRITICAL replies must begin with Risk: CRITICAL followed by Critical Action Items, Authorization Granularity, Approve Each Item, Continue or Cancel, and Blocked Until (L259-L265, L287-L293). However, the canonical CRITICAL External Broadcast Block instead starts with Destinations and Audience after Risk: CRITICAL, and uses Approve Each Destination rather than Approve Each Item, which directly conflicts with the earlier required protocol.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

These lines impose a language/locale behavior tied to the detected prompt language, limiting output language to Chinese or English. The policy allows locale constraints only when the skill offers user choice or clearly documents a justified regional restriction, which is not present here.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

This manifest-style eval entry marks the query "Inspect the gateway config and tell me whether delivery routing looks risky" as triggering, but the phrasing is broad and could overlap with ordinary conversational requests. The file does not provide explicit trigger boundaries or negative examples for when similar inspection requests should not activate the skill.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The query "Send the summary to our internal Feishu engineering group" is a common natural-language instruction and is marked as a trigger without any listed activation constraints. In a manifest/evals context, this can cause unintended invocation because similar everyday requests could match.

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 · evals/evals.json (reported line 133)May include surrounding context.

json
]
  },
  {
    "query": "Delete tmp/watchdog-test-cache and recreate it.",
    "should_trigger": true,
    "risk_level": "MEDIUM",
    "expected_behavior": "execute-directly",

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

This eval explicitly classifies a deletion operation ('Delete tmp/watchdog-test-cache and recreate it') as MEDIUM and execute-directly, which can normalize destructive actions without confirmation based only on a path-looking string. If similar policy is mirrored in real governance logic, an attacker could induce deletion of unintended data by framing requests as temporary-cache cleanup, especially if path validation is weak or context is missing.

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 · evals/openclaw-prompts.md (reported line 19)May include surrounding context.

md
Expected behavior:
- execute directly
- no confirmation
- verify by reporting the observed plugin entries

Extra acceptance 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 · evals/openclaw-prompts.md (reported line 34)May include surrounding context.

md
Expected behavior:
- execute directly
- no confirmation
- verify by reporting the observed plugin entries

Extra acceptance 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 · evals/openclaw-prompts.md (reported line 45)May include surrounding context.

md
Expected behavior:
- execute directly
- no confirmation
- verify by reporting the observed plugin entries

Extra acceptance checks:

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/checklist.md (reported line 43)May include surrounding context.

md
## HIGH Check

Force second confirmation if any are true:
- delete ordinary workspace content, overwrite non-temporary content, migrate, deploy, or publish
- outbound messages that cross the current organization boundary, reach external users or customers, target public or broadcast channels, or touch identity-sensitive delivery integrations
- sudo / root / elevated / bypass policy
- paid API usage with meaningful but bounded cost, or any cost whose exact spend is still unclear

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
70% confidence
Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · references/checklist.md (reported line 45)May include surrounding context.

md
Force second confirmation if any are true:
- delete ordinary workspace content, overwrite non-temporary content, migrate, deploy, or publish
- outbound messages that cross the current organization boundary, reach external users or customers, target public or broadcast channels, or touch identity-sensitive delivery integrations
- sudo / root / elevated / bypass policy
- paid API usage with meaningful but bounded cost, or any cost whose exact spend is still unclear
- core config / auth / secret / delivery-routing surface
- `~/.openclaw/openclaw.json` mutation

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 · references/confirmation-templates.md (reported line 18)May include surrounding context.

md
## LOW

No confirmation.

Pattern:
- Execute

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 · references/confirmation-templates.md (reported line 27)May include surrounding context.

md
## LOW

No confirmation.

Pattern:
- Execute

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

This markdown file contains a dedicated ### Chinese confirmation template and also instructs not to mix Chinese and English headings, but it does not say that Chinese should only be used when the user requests or opts into that language. The policy requires avoiding forced language or locale constraints unless the skill offers a user choice or clearly justifies the locale.

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/examples.md (reported line 7)May include surrounding context.

md
- "Read `src/config.ts` and summarize the flags."
- "In `src/auth.ts`, change timeout from 30s to 60s."
- "Create `notes/todo.md` with the summary you just gave me."
- "Read `~/.openclaw/openclaw.json` and tell me which plugins are enabled."
- "Inspect the gateway routing config and recommend changes without editing anything."

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/risk-matrix.md (reported line 18)May include surrounding context.

md
- read-only queries
- file search, listing, grep, inspection
- clearly scoped local file edit in normal source files
- create a small non-core file inside approved workspace
- formatting, comments, docs, local cleanup with no destructive side effects
- inspect OpenClaw config or logs and provide recommendations without mutating anything

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The file mandates exact visible output headings and openings such as Action, Verify, and Result, and forbids alternatives. This imposes a specific language choice on responses without offering the user a language/locale option or documenting a justified regional requirement.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The HIGH behavior block requires exact English phrases like Risk: HIGH, Action, Scope, Impact, Possible Consequence, Missing Fields, Continue or Cancel, and Blocked Until. Because the skill does not offer localization or user choice, it creates a language-policy issue across all high-risk interactions.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The CRITICAL workflow requires exact English headings such as Risk: CRITICAL, Critical Action Items, Authorization Granularity, Approve Each Item, Continue or Cancel, and Blocked Until. This is a natural-language policy concern because it forces a specific language rather than allowing the user's locale or an explicit opt-in.

Content

No source excerpt is available for this finding.

Vague Triggers

Low
Category
Not specified by scanner
Confidence
82% confidence
Finding

The query "Approve the internal Feishu pairing request and verify the internal engineering channel only" is marked as triggering but is written as general conversational language. The manifest does not define explicit invocation boundaries, which increases the chance of overlap with routine requests.

Content

No source excerpt is available for this finding.