Back to skill

Security audit

Multi-Agent Dev Team

Security checks for vulnerabilities and agentic risk

Overview

This skill is a coherent software-development workflow, but it gives agents broad file, shell, Git, dependency-installation, and possible publishing authority without enough upfront scoping or approval controls.

Install only in a dedicated project workspace without production secrets. Review planned commands, dependency additions, generated diffs, commits, remotes, and deployment targets before allowing pushes or deployments. Prefer pinned installer and scaffolding versions, lockfiles, and explicit confirmation before any GitHub or Vercel publication.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • 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)

T08 · Insecure Dependencies

Warning
Location
SKILL.md:29
Finding
Unpinned npm Packages Can Execute Mutable Third-Party Code## Vulnerability Details **File Locations**: - `SKILL.md:29-31` - `SKILL.md:335-337` - `README.md:21-23` - `README.md:202-204` - `agents/dev-agent/SOUL.md:197-204` - `agents/dev-agent/SOUL.md:213-218` **Vulnerability Type**: Unpinned package execution and supply-chain exposure **Risk Level**: Medium ### Vulnerable Code `SKILL.md:29-31` and `SKILL.md:335-337`: ```bash npx clawhub install multi-agent-dev-team ``` `README.md:21-23` and `README.md:202-204`: ```bash npx clawhub install multi-agent-dev-team ``` `agents/dev-agent/SOUL.md:197-204`: ```bash # Next.js npx create-next-app@latest project-name # Node.js npm init -y npm install <packages> ``` `agents/dev-agent/SOUL.md:213-218`: ```bash npm install # Install dependencies npm run dev # Start dev server npm run build # Production build npm test # Run tests ``` ### Technical Analysis The documented installation workflow uses `npx` without pinning `clawhub` to an exact, previously reviewed version. The Dev agent is also instructed to execute `create-next-app@latest` and install unspecified packages. `npx` can retrieve a package from the configured npm registry and execute its entry point immediately. The `@latest` tag is mutable and may resolve to different code over time. Similarly, `npm install <packages>` permits the agent to select and install dependencies that were not included in this audit. npm installation may execute package lifecycle hooks such as `preinstall`, `install`, and `postinstall`, giving dependency code execution rights under the agent's operating-system account. The repository itself contains no declared dependencies, package scripts, lockfile, integrity metadata, or package allowlist that would constrain the effective code executed by these commands. Therefore, the reviewed project content does not fully determine the code that will run during installation or generated-project setup. Potential exploitation methods include: 1. Co ...[truncated 2036 chars]
Remediation
## Remediation Suggestions 1. Pin all directly executed tools to reviewed, exact versions: ```bash npx --yes clawhub@<reviewed-exact-version> install multi-agent-dev-team npx --yes create-next-app@<reviewed-exact-version> project-name ``` 2. Remove `@latest` from agent instructions. Upgrade versions only through an explicit review process. 3. Require generated projects to commit and enforce a lockfile, using deterministic installation commands such as: ```bash npm ci ``` 4. Maintain an allowlist of approved package names, versions, registries, and integrity hashes. Do not permit the Dev agent to infer arbitrary dependency names without review. 5. Configure npm to use an explicitly trusted registry and enable provenance or signature verification where supported. 6. Disable lifecycle scripts when they are unnecessary: ```bash npm ci --ignore-scripts ``` If lifecycle scripts are required, inspect the relevant packages and scripts before allowing them to execute. 7. Run package installation in an isolated, least-privileged environment without production secrets, broad filesystem access, inherited SSH agents, or unnecessary network access. 8. Add dependency scanning and lockfile review to the PM agent's acceptance criteria. Reject unexpected package additions, registry changes, Git dependencies, local path dependencies, and install-script changes. 9. Document that credentials should not be exposed to package installation processes and that generated code must undergo human review before being pushed or deployed.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (21)

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The overview states that the Dev agent implements code and commits to Git, but it does not warn users that invoking the skill may modify local files and create repository history. In an agentic coding context, omission of these side-effect warnings can lead users to run the skill in sensitive directories or on existing repositories without realizing the consequences.

Rp1

Medium
Category
MCP Rug Pull
Confidence
90% confidence
Finding
The README instructs users to run `npx clawhub install multi-agent-dev-team` without pinning a specific package version. This allows whatever version is current at execution time to be fetched and run, which creates a supply-chain risk if the package or a dependency is compromised or a malicious update is published.

Missing User Warnings

Medium
Confidence
93% confidence
Finding
The quick-start section tells users to launch the PM agent and 'simply describe what you want to build' without disclosing that the workflow may generate a full project and perform Git operations. This is risky because quick-start instructions are the most likely path users follow, and the lack of safety notice reduces informed consent for workspace-modifying behavior.

Missing User Warnings

Medium
Confidence
96% confidence
Finding
The README advertises that the output includes a Git repository initialized and production-ready code, but it does not pair this with a warning about repository-changing actions. In the context of an autonomous multi-agent development skill, normalizing automatic repository mutation without caution increases the chance of unintended commits, overwrites, or confusion about what the agent changed.

Rp1

Medium
Category
MCP Rug Pull
Confidence
90% confidence
Finding
The repeated use of unpinned `npx clawhub` in the closing install example has the same supply-chain risk as the earlier occurrence. Because `npx` may download and execute the latest available package, users can be exposed to unintended or malicious code changes over time.

Rp1

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding
The installation command uses `npx clawhub install multi-agent-dev-team` without a pinned version, which allows whatever package/version is current at execution time to be fetched and run. In a skill that orchestrates local development agents and may modify files, unpinned installer execution increases supply-chain risk and could expose users to malicious or unexpectedly changed code.

Missing User Warnings

Medium
Confidence
93% confidence
Finding
The skill describes that the Dev agent 'implements code', 'tests functionality', and 'commits to Git' but does not present a clear, upfront safety warning that local files may be modified automatically and commits may be created. In the context of a multi-agent development workflow, users may authorize the skill without realizing it can change their workspace state, overwrite files, or create unintended repository history.

Rp1

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding
The repeated `npx clawhub install multi-agent-dev-team` command again references an unpinned package/version, creating the same supply-chain exposure at the call-to-action section. Repetition increases the chance users will copy-paste and execute it without scrutiny, which is risky because `npx` may download and run code immediately.

Missing User Warnings

Medium
Confidence
89% confidence
Finding
The instruction to push code to GitHub when credentials are available authorizes external data transmission without requiring review of what is being sent. In an agent setting, this can expose proprietary code, secrets accidentally written to the repo, or other sensitive workspace contents to a third-party service.

Missing User Warnings

Medium
Confidence
84% confidence
Finding
Listing `Write` and `exec` as normal tools to use, without clear safety constraints, encourages workspace and system modifications by default. In a multi-agent workflow this increases the chance of destructive commands, unintended file changes, dependency installation, or execution of untrusted project scripts.

Rp1

Medium
Category
MCP Rug Pull
Confidence
87% confidence
Finding
The skill recommends running `npx create-next-app@latest`, which fetches and executes the latest package version at runtime without pinning. That creates a supply-chain risk: future upstream changes or a compromised release could cause the agent to execute unexpected code during project setup.

Description-Behavior Mismatch

Medium
Confidence
96% confidence
Finding
The skill defines the PM agent as an orchestrator, yet grants it direct file write access and command execution. This violates separation of duties and least privilege: if prompted or subverted, the PM agent can make unreviewed code or environment changes instead of only coordinating work, increasing the attack surface and bypassing the intended Dev-agent handoff.

Context-Inappropriate Capability

Medium
Confidence
94% confidence
Finding
Granting shell execution to a project-management agent is unnecessary for its stated role and enables arbitrary command execution in the workspace or host context. An attacker could exploit prompt injection or task confusion to run destructive, exfiltrating, or persistence-establishing commands under the PM agent's authority.

Missing User Warnings

Medium
Confidence
96% confidence
Finding
Advertising `exec` without warning that shell/system commands may run conceals a significant capability from the user. In a multi-agent workflow, this is especially risky because users may delegate high-level planning tasks without realizing the PM agent can execute commands that affect the filesystem, network, or runtime environment.

Intent-Code Divergence

Medium
Confidence
95% confidence
Finding
The document explicitly says the PM agent is 'not the implementer,' but its toolset allows direct implementation actions. This mismatch is dangerous because operators may rely on the textual role description while overlooking the effective privileges, leading to overtrust and making misuse or prompt-induced action more likely.

Description-Behavior Mismatch

Medium
Confidence
95% confidence
Finding
The example materially expands the advertised 2-agent workflow into a broader system involving QA and DevOps agents, including testing and deployment actions. This mismatch can mislead users and reviewers about the skill's actual capabilities and trust boundaries, increasing the risk of unexpected autonomous actions being authorized.

Description-Behavior Mismatch

Medium
Confidence
93% confidence
Finding
The documentation presents a software development workflow but the example includes publishing code to GitHub and deploying a live site to Vercel, which are external side effects beyond simple code generation. Users may reasonably interpret the skill as local-only development assistance and unknowingly permit public release of code or artifacts.

Context-Inappropriate Capability

Medium
Confidence
91% confidence
Finding
External deployment and repository publication are high-impact capabilities because they can exfiltrate source code, expose unfinished applications publicly, and consume third-party credentials or quotas. Presenting these actions as routine without tying them to the stated skill purpose weakens least-privilege expectations and broadens the attack surface.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The deployment guidance lacks prominent warnings that using the workflow will publish code and a live application to third-party services. Without informed consent, users may expose proprietary code, test content, environment assumptions, or brand assets to public infrastructure unintentionally.

External Transmission

Medium
Category
Data Exfiltration
Content
**API:**
- Use GitHub's public API (no auth required)
- Endpoint: `https://api.github.com/users/{username}`
```

## Tips for Writing Good Specs
Confidence
50% confidence
Finding
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.

Missing User Warnings

Low
Confidence
90% confidence
Finding
The skill instructs the PM agent to write to project and memory files without clearly warning the user that files may be modified. This weakens informed consent and can cause unintended alteration of repository state, planning artifacts, or stored memory, especially when users expect a coordination-only role.

Static analysis

No suspicious patterns detected.