Back to skill

Security audit

Clawd Cursor 0.6.0

Security checks for vulnerabilities and agentic risk

Overview

This skill is a powerful desktop-control agent, and its reviewed artifact leaves too much authority to mutable remote code and agent-started local services.

Install only if you are comfortable with a local service that can see and operate your desktop and browser. Prefer a version pinned to a reviewed commit with locked dependencies, keep Chrome remote debugging disabled unless intentionally needed, use a local provider such as Ollama for private screens, and require explicit confirmation before the service is started or used on sensitive apps.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • 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
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (1)

T03 · Remote Payload Retrieval and Execution

Error
Location
SKILL.md:24
Finding

Unpinned Remote Code and Dependency Retrieval with Local Execution

Content
View full analysis
Remediation
View remediation
test "$(git rev-parse HEAD)" = "" ``` 2. **Bundle auditable implementation source** - Include the scripts and source needed by the Skill in the published artifact. - Ensure the reviewed package corresponds exactly to the code executed after installation. 3. **Lock and verify npm dependencies** - Commit a reviewed `package-lock.json`. - Use `npm ci` instead of `npm install` to enforce the lockfile. - Review transitive dependencies and retain npm integrity hashes. - Use an approved registry and prevent dependency substitution. 4. **Restrict lifecycle-script execution** - Use `npm ci --ignore-scripts` where possible. - Explicitly review and separately invoke any lifecycle scripts that are genuinely required. - Perform installation and building in a sandbox with minimal filesystem, credential, and network access. 5. **Avoid unpinned `npx` execution** - Invoke a verified local binary from the locked dependency tree. - If package retrieval is unavoidable, specify an exact reviewed version and verify package provenance and integrity. - Configure `npx` so it fails rather than downloading an absent executable unexpectedly. 6. **Apply runtime least privilege** - Run the desktop service as a dedicated, non-administrative account where practical. - Do not expose unrelated environment variables or credentials to the installation process. - Restrict outbound connections to the explicitly selected AI provider. - Verify that ports `3847` and `9222` are bound only to l ...[truncated 352 chars]
Vulnerability Patterns
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (8)

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The description is extremely broad (control any app, if you can click it, your agent can too), which increases the chance the skill is invoked for many unrelated requests. In a desktop-control skill, over-broad routing is security-relevant because activation grants access to sensitive on-screen data and powerful local actions.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
84% confidence
Finding

The install flow invokes npx clawd-cursor doctor without pinning an explicit package version, so execution depends on whatever version resolves at runtime. Even though the repo itself is cloned at install time, unpinned npx usage still introduces supply-chain risk if the package registry serves a newer or compromised release than expected.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The guidance says to use the skill when the user asks for desktop interaction or when no other tool can accomplish the task, which is ambiguous and may encourage opportunistic use. Because this skill can read screens and operate applications, ambiguous activation materially raises the risk of unnecessary exposure or unintended actions.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill documents direct use of Chrome's remote debugging interface, which grants powerful browser control and page data access outside the stated Clawd Cursor REST API. This meaningfully broadens agent capability: CDP can inspect DOM content, interact with authenticated sessions, and bypass the skill's own approval or logging model if Chrome is exposed with remote debugging enabled.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
94% confidence
Finding

The instruction to start the server yourself and specifically 'don't ask the user' promotes autonomous local process execution without explicit consent. In the context of a desktop-control skill, encouraging silent startup lowers user awareness, can expand the local attack surface, and normalizes unsupervised enablement of a powerful automation service.

Content

Scanner excerpt · SKILL.md (reported line 124)May include surrounding context.

Expected: {"status":"ok","version":"0.6.0"}

If connection refused — start it yourself (don't ask the user):

powershell
# Find the skill directory and start the server
Start-Process -FilePath "node" -ArgumentList "dist/index.js","start" -WorkingDirectory "<clawd-cursor-directory>" -WindowStyle Hidden

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
83% confidence
Finding

The privacy section says that, with cloud providers, screenshots/text go only to the chosen provider API and not elsewhere. But the same file documents an alternate fast path where agents directly connect to Chrome's debugging port and read page text, titles, URLs, and interact with elements locally, meaning data flow is not limited to the provider path described there.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
84% confidence
Finding

The setup section again instructs use of npx clawd-cursor doctor without a pinned version, which can execute code fetched dynamically from the package registry. This creates a supply-chain exposure and makes behavior non-reproducible across installations.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
79% confidence
Finding

The performance tooling references npx ts-node perf/perf-test.ts without pinning a version, allowing execution of whichever ts-node version is resolved in the environment or registry. While this appears intended for benchmarking, it still expands the trusted code boundary and introduces avoidable supply-chain risk.

Content

No source excerpt is available for this finding.

Static analysis

No suspicious patterns detected.