Back to skill

Security audit

clawlite-mark

Security checks for vulnerabilities and agentic risk

Overview

This skill is mostly a disclosed Facebook automation helper, but it asks the agent to use a separate private workspace persona file while posting publicly from a persistent Facebook session.

Review before installing. Use a dedicated Facebook profile, prefer draft or per-reply approval mode, and remove or replace the cross-workspace `/Users/m1/.openclaw/workspace-elon/SOUL.md` reference with a packaged style guide before allowing public auto-replies.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (1)

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:36
Finding

Cross-Workspace Persona File Access Violates Least Privilege

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, line 36
Vulnerability Type: Cross-workspace access to unrelated instructions and private context
Risk Level: Medium

Complete Snippet:

markdown
**Reference:** See Elon's voice guidelines in `/Users/m1/.openclaw/workspace-elon/SOUL.md` for writing style.

Technical Analysis

The skill explicitly directs the agent to read an absolute path in another agent's workspace. Access to workspace-elon/SOUL.md is not necessary to perform the stated Facebook posting and comment-management functions. This violates least privilege and crosses the skill's expected workspace boundary.

A persona or instruction file may contain private context, operational rules, or other workspace-specific information. It is also an externally mutable input: anyone able to modify that file could influence the skill's generated public content without changing the audited package. The reference therefore creates both a confidentiality concern and an indirect instruction-injection channel.

Attack Path

  1. A user invokes the Facebook automation skill.
  2. The agent follows the style-reference instruction in SKILL.md.
  3. The agent reads /Users/m1/.openclaw/workspace-elon/SOUL.md, assuming filesystem permissions permit access.
  4. Private or attacker-controlled instructions in that file enter the reply-generation context.
  5. The agent may incorporate those instructions or sensitive context into Facebook posts or replies.
  6. In automatic approval mode, affected content may be published without per-message human review.

Impact Assessment

Successful exploitation is limited to the privileges already held by the agent process; this instruction does not independently bypass operating-system access controls. However, if the process can read the referenced path, it gains access to another workspace's persona and instruction state beyond the Facebook task's legitimate requirement ...[truncated 256 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove the absolute cross-workspace reference from SKILL.md.
  2. Package a minimal, reviewed style guide within this skill's own directory.
  3. Restrict filesystem access to the project directory and explicitly approved output locations.
  4. Require explicit user authorization before reading files outside the skill workspace.
  5. Treat all external persona and instruction files as untrusted data rather than executable instructions.
  6. Apply provenance controls and content filtering before external context affects public posts.
  7. Require draft-and-approve mode whenever external context is used to generate Facebook content.
  8. Log authorized cross-workspace reads without recording sensitive file contents.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (4)

Natural-Language Policy Violations

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The instruction 'ENGLISH ONLY' and 'No exceptions' forces a specific language for all posts, comments, and replies. This is a natural-language policy violation because it imposes a locale/language restriction without offering user choice, opt-in, or a documented justification for the constraint.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
87% confidence
Finding

The skill is designed to operate against a persistent logged-in Facebook browser profile and to perform posting and reply actions, which implies reuse of an authenticated session. Persistent session automation raises the risk of unauthorized actions if the skill is triggered unexpectedly, if the profile is shared, or if other instructions in the environment steer the browser toward unintended operations.

Content

Scanner excerpt · SKILL.md (reported line 20)May include surrounding context.

md
## What Mark Does

1. **Post to Facebook** — Create posts using the existing posting infrastructure
2. **Read Comments** — Fetch all comments on a specific post or recent posts
3. **Generate Replies** — Use AI to generate contextually appropriate replies
4. **Post Replies** — Automatically reply to comments

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill explicitly supports creating Facebook posts and automatically replying to comments on a live account, but the core capability description does not prominently warn that these actions have external side effects on a real third-party platform. That increases the risk of unintended publication, accidental spam, or unauthorized brand communications if invoked casually or mis-scoped by an agent.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The usage examples include commands that directly publish content and reply to each comment, yet they are presented without an adjacent warning that these actions are irreversible or externally visible. Example-driven invocation is especially risky because users and agents often copy examples verbatim, which can trigger mass posting or engagement without adequate review.

Content

No source excerpt is available for this finding.