Back to skill

Security audit

Temp Agent Browser

Security checks for vulnerabilities and agentic risk

Overview

This browser automation skill is coherent, but it needs review because it documents unpinned third-party installation and persistent authenticated browser state without enough safety guidance.

Install only if you are comfortable reviewing and trusting the upstream agent-browser package or pinning it yourself. Avoid using saved auth state for high-privilege accounts, protect any auth.json, screenshot, PDF, video, or trace files like sensitive data, and do not pass real passwords as plain command arguments unless you accept the exposure risk.

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:17
Finding
Unpinned Third-Party Installation and Source Build## Vulnerability Details **File Location**: `SKILL.md:17-19`, `SKILL.md:24-27`, and `CONTRIBUTING.md:18` **Vulnerability Type**: Supply-chain exposure through mutable, unverified dependencies **Risk Level**: Medium ### Vulnerable Code `SKILL.md:17-19`: ```bash npm install -g agent-browser agent-browser install agent-browser install --with-deps ``` `SKILL.md:24-27`: ```bash git clone https://github.com/vercel-labs/agent-browser cd agent-browser pnpm install pnpm build ``` `CONTRIBUTING.md:18`: ```bash npm install -g agent-browser@latest ``` ### Technical Analysis The documented installation procedures retrieve and execute mutable third-party content without pinning it to a reviewed package version, source commit, or verified integrity digest. The npm commands install either the registry's currently selected release or explicitly use the mutable `latest` tag. The source-build procedure clones the upstream repository's current default branch and then resolves dependencies without a project-provided lockfile or other reproducibility control. npm package installation may execute package lifecycle scripts. Likewise, `pnpm install`, `pnpm build`, and the subsequent browser installation commands can execute code supplied by the upstream project or its transitive dependencies. Consequently, the code executed by users can differ from the code that was reviewed when this Skill was published. The `agent-browser install --with-deps` operation may also install operating-system dependencies and can require elevated privileges, increasing the potential effect of a compromised upstream installer. The audited files do not themselves contain a malicious payload, and exploitation depends on compromise or malicious modification of an upstream package, repository, release tag, or dependency. ### Attack Path 1. An attacker compromises the `agent-browser` npm package, its maintainer account, the upstream repository, or ...[truncated 1378 chars]
Remediation
## Remediation Suggestions 1. Pin the npm dependency to a specific, reviewed version rather than relying on an implicit current release or `@latest`: ```bash npm install -g agent-browser@<reviewed-version> ``` 2. Document the expected npm integrity digest and provide a verification procedure before installation. 3. For source installations, pin the repository to a full reviewed commit hash: ```bash git clone https://github.com/vercel-labs/agent-browser cd agent-browser git checkout --detach <reviewed-full-commit-hash> ``` 4. Provide and enforce a reviewed lockfile so transitive dependency resolution is reproducible. Use frozen-lockfile installation where supported: ```bash pnpm install --frozen-lockfile ``` 5. Prefer signed releases or commits and instruct users to verify signatures before building. 6. Remove the recommendation to install `@latest` during troubleshooting. Instead, identify a supported, security-reviewed version and update it through a controlled review process. 7. Clearly separate ordinary user-level setup from operations that may require administrative access. Explain exactly what `--with-deps` changes and advise users to inspect the generated package-manager operations before granting elevation. 8. Run installation and builds in a restricted environment where practical, with minimal filesystem access, no unnecessary secrets in the environment, and no administrative privileges.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
  • Rogue AgentSelf-Modification, Session Persistence
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (7)

Self-Modification

High
Category
Rogue Agent
Content
## Adding New Commands to the Skill

Update SKILL.md when the upstream CLI adds new commands.
- Keep the Installation section
- Add new commands in the correct category
- Include usage examples
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.

Context Window Stuffing

Medium
Category
Memory Poisoning
Content
agent-browser open <url>        # Navigate to page
agent-browser snapshot -i       # Get interactive elements with refs
agent-browser click @e1         # Click element by ref
agent-browser fill @e2 "text"   # Fill input by ref
agent-browser close             # Close browser
```
Confidence
85% confidence
Finding
Skill attempts to fill the context window with filler content, displacing legitimate instructions and safety constraints. This can degrade agent performance or bypass safety boundaries.

Missing User Warnings

Medium
Confidence
98% confidence
Finding
The documented command `agent-browser set credentials user pass` encourages passing secrets directly on the command line, where they may be exposed via shell history, process listings, audit logs, or agent transcripts. In an AI-agent environment, this materially increases the chance of credential disclosure beyond the immediate terminal session.

Missing User Warnings

Medium
Confidence
95% confidence
Finding
The skill documents saving and reloading browser session state (e.g. auth.json) without warning that the file may contain cookies, tokens, and authenticated session material. In an agent context, users may persist privileged sessions to disk and later expose them through reuse, leakage, or accidental inclusion in logs or artifacts.

Session Persistence

Medium
Category
Rogue Agent
Content
```bash
agent-browser state save auth.json    # Save session state
agent-browser state load auth.json    # Load saved state
```

## Example: Form submission
Confidence
94% confidence
Finding
The `state load auth.json` command explicitly enables reuse of persisted authenticated browser state. Without safeguards or warnings, an agent may load stale, overprivileged, or attacker-provided session artifacts, resulting in account takeover, unauthorized actions, or cross-task credential bleed.

Session Persistence

Medium
Category
Rogue Agent
Content
agent-browser wait --url "/dashboard"
agent-browser state save auth.json

# Later sessions: load saved state
agent-browser state load auth.json
agent-browser open https://app.example.com/dashboard
```
Confidence
94% confidence
Finding
The example workflow normalizes loading previously saved authentication state for later sessions, but does not warn that the state file can act as a bearer credential. In agent workflows this increases the risk of persistent session reuse, accidental sharing, or use of an authentication artifact outside its intended boundary.

Missing User Warnings

Low
Confidence
79% confidence
Finding
The skill describes saving screenshots, PDFs, videos, traces, and session state to local files, but does not include any general warning that these commands write potentially sensitive browser contents to disk. For a browser automation skill, such artifacts can contain private page data, credentials, or user content.

Static analysis

No suspicious patterns detected.