Agent Browser Local

• Outcome: Review
• Latest audit: Sep 12, 2026 · 5:14 AM UTC
• Version: 1.0.0

Overview

This browser automation skill is coherent and not malicious, but it should be reviewed because it installs a mutable third-party CLI and documents persistent authenticated browser state without adequate safety guidance.

Review the installer before use, prefer a pinned and trusted agent-browser version, avoid running installer commands with elevated privileges, and treat saved browser state, screenshots, traces, PDFs, and recordings as sensitive. Use separate per-task sessions and delete auth.json or other captured artifacts when no longer needed, especially for important accounts.

A.I.G

Vulnerability Patterns

• Insecure Dependencies: Introduces malicious components through unsafe dependency sources
• Skill Instruction Hijacking: Alters the agent's session goals or safety constraints when the skill loads
• Agent Memory Poisoning: Writes attacker-controlled rules into memory that affect later sessions
• Remote Payload Retrieval and Execution: Fetches external code whose behavior can change after review
• Embedded Malicious Code: Ships malicious scripts inside the skill and executes them locally

Findings (1)

T08 · Insecure Dependencies

Warning

• Location: SKILL.md:17
• Finding: Unpinned Third-Party Package and Source Installation## Vulnerability Details **File Locations**: - `SKILL.md:17-26` - `CONTRIBUTING.md:23-26` **Vulnerability Type**: Unpinned and mutable third-party dependencies **Risk Level**: Medium ### Vulnerable Code `SKILL.md:17-26` ```bash npm install -g agent-browser agent-browser install agent-browser install --with-deps ``` ```bash git clone https://github.com/vercel-labs/agent-browser cd agent-browser pnpm install pnpm build agent-browser install ``` `CONTRIBUTING.md:23-26` ```bash 1. Install the latest version ```bash npm install -g agent-browser@latest ``` ``` ### Technical Analysis The installation instructions retrieve and execute mutable third-party content without pinning an audited package version, source commit, or dependency integrity value. `npm install -g agent-browser` resolves the package version according to the npm registry's current state, while `agent-browser@latest` explicitly follows a mutable distribution tag. Consequently, the installed artifact can change after this Skill has been reviewed. npm package installation may also execute lifecycle scripts with the privileges of the invoking user. The source installation procedure clones the repository's default branch rather than a specific audited commit or signed release. It then runs `pnpm install` and `pnpm build`, allowing the current upstream source and its transitive dependency graph to execute build or lifecycle logic. No lockfile, frozen dependency resolution, checksum, signature, or provenance verification is required by the instructions. The subsequent `agent-browser install --with-deps` command may install additional browser or system dependencies, increasing the amount of externally sourced code and potentially requiring elevated permissions depending on the host configuration. ### Attack Path 1. An attacker compromises the npm package, its mutable `latest` release, the upstream repository, a maintainer ac ...[truncated 1666 chars]
• Remediation: ## Remediation Suggestions 1. Pin `agent-browser` to a specifically reviewed version rather than using an implicit version or `@latest`, for example: ```bash npm install --global agent-browser@<audited-version> ``` 2. Record and verify the expected npm package integrity digest and package provenance before installation. Document the trusted publisher and official package identity to reduce dependency-confusion and package-substitution risk. 3. For source installation, check out a specific audited commit or signed release: ```bash git clone https://github.com/vercel-labs/agent-browser cd agent-browser git checkout --detach <audited-commit-sha> ``` Verify the commit or release signature where upstream signing is available. 4. Require a committed lockfile and use frozen dependency resolution, such as: ```bash pnpm install --frozen-lockfile ``` Review dependency changes before updating the pinned source revision or package version. 5. Avoid global installation where possible. Prefer an isolated project environment, container, or restricted execution account with access only to the files and network destinations required for browser automation. 6. Do not run npm, pnpm, build scripts, or `agent-browser install --with-deps` with administrative privileges unless explicitly necessary. Separate operating-system dependency installation from package installation and review the exact system changes first. 7. Document expected lifecycle and installer behavior. Where practical, initially inspect packages with lifecycle scripts disabled, review the scripts, and only then permit required installation actions. 8. Replace the recommendation in `CONTRIBUTING.md` to install `@latest` with the same audited version used by the Skill, and update that version only after a new security review.

SkillSpector

Vulnerability Patterns

• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Memory Poisoning: Persistent Context Injection, Context Window Stuffing, Memory Manipulation
• Rogue Agent: Self-Modification, Session Persistence
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Privilege Escalation: Excessive Permissions, Sudo/Root Execution, Credential Access

Findings (6)

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: 89% confidence
• Finding: The skill explicitly demonstrates saving and reloading authenticated browser state (`auth.json`) but provides no warning that this file may contain reusable session cookies, tokens, or other authentication material. In an agent context, this can lead to accidental credential persistence, reuse across tasks, or leakage if the file is stored insecurely, shared, or committed.

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: 93% confidence
• Finding: `state save` and `state load` enable persistence of browser session data across runs, which can include authenticated cookies and storage. In an agent-operated environment, persisted session artifacts can be reused by unintended tasks or users if not protected, effectively turning a convenience feature into credential/session replay risk.

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 normalizes reusing previously saved authenticated state without discussing trust boundaries or expiry, which may encourage insecure long-lived session reuse. If such files are copied, exposed, or left on disk, an attacker or unrelated workflow could inherit authenticated access without needing credentials.

Missing User Warnings

Low

• Confidence: 84% confidence
• Finding: The documentation describes file-writing operations such as screenshots, PDFs, recordings, traces, state files, and uploads without warning that these actions create or consume local files that may contain sensitive page content or overwrite existing data. In an automation skill, agents may perform these actions non-interactively, increasing the risk of unintended local data exposure or persistence.

Static analysis

No suspicious patterns detected.