Back to skill

Security audit

Tabbit Agent Browser

Security checks for vulnerabilities and agentic risk

Overview

The skill is coherent for connecting an agent to Tabbit, but it exposes high-privilege browser control and can automatically run an unpinned npm package.

Install only if you trust the local Tabbit session and the agent-browser executable being run. Prefer a preinstalled, trusted, pinned agent-browser binary, avoid exposing DevTools debugging beyond localhost, and disable Tabbit remote debugging when finished.

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
scripts/run_agent_browser_on_tabbit.py:58
Finding

Unpinned Third-Party Package Download and Execution

Content
View full analysis

Vulnerability Details

File Location: scripts/run_agent_browser_on_tabbit.py, lines 58–59
Vulnerability Type: Unpinned dependency execution through npx
Risk Level: Medium

python
if shutil.which("npx"):
    return ["npx", "--yes", "agent-browser"]

Technical Analysis

When a locally installed agent-browser executable is unavailable, the wrapper automatically falls back to npx --yes agent-browser. This command may retrieve and execute the package currently published under that name without an exact version, lockfile, integrity constraint, or interactive approval.

Consequently, the effective code executed by the audited skill can change after review. The trust boundary extends to the package registry, the package publisher, and the package's transitive dependency chain. The --yes option removes the normal opportunity for the user to review or reject package installation.

Attack Path

  1. The victim invokes scripts/run_agent_browser_on_tabbit.py.
  2. No agent-browser executable is available on PATH, but npx is installed.
  3. The wrapper selects npx --yes agent-browser automatically.
  4. npx resolves the unpinned package through its configured registry and may download it and its dependencies.
  5. A compromised or malicious package release executes with the privileges of the user running the wrapper.
  6. The package receives the Tabbit CDP WebSocket endpoint and browser-operation arguments, enabling it to interact with the connected browser session.

Exploitation depends on compromise or malicious control of the resolved package, its publisher, the configured registry, or its dependency chain.

Impact Assessment

Malicious package code could execute arbitrary commands with the invoking user's operating-system privileges and access files, environment variables, and network resources available to that user. Because the wrapper supplies a live Chrome DevTools Protocol endpoint, the malicious code could also control Ta ...[truncated 330 chars]

Remediation
View remediation

Remediation Suggestions

  • Remove the automatic npx --yes fallback and require a preinstalled, explicitly trusted agent-browser executable by default.
  • If npx support is required, pin an audited exact package version, for example agent-browser@<exact-version>, rather than resolving the latest release.
  • Use a lockfile and registry-supported integrity verification to pin both the direct package and its transitive dependencies.
  • Require explicit user confirmation before any package is downloaded or executed; do not use --yes for an implicit fallback.
  • Restrict package resolution to an approved registry and document the expected package publisher and version.
  • Prefer installation during a controlled setup phase instead of downloading executable code during browser automation.
  • Verify the resolved executable or package artifact against an approved checksum or signature before launching it.
  • Fail closed with a clear installation message when no trusted binary is available.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Behavioral ASTexec() Call, eval() Call, Dynamic Import
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
Findings (7)

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
93% confidence
Finding

The skill explicitly instructs the agent to read local files under the user's home directory and run helper scripts/wrappers, which implies file-read and shell capabilities, yet it declares no explicit tool scope or allowed-tools boundary. Without a declared permission boundary, an orchestrator may over-grant capabilities or users may be unaware that invoking this skill can trigger local file access and command execution.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The activation rule says to prefer this skill whenever a request is about Tabbit or contains several phrases, but it does not clearly bound when the skill should not activate. Over-broad routing can cause the agent to invoke a skill that reads local DevToolsActivePort files and launches shell-based wrappers in situations where the user only mentioned Tabbit conversationally, increasing unintended access to local browser-debugging endpoints.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The user-facing short description and default prompt are entirely in Chinese, with no indication that another language is available or that the user can choose their preferred locale. This creates a language-policy issue because the skill effectively mandates one language rather than offering an opt-in or choice.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The setup explicitly instructs users to enable remote debugging and connect via the DevTools WebSocket endpoint, which grants powerful browser control. Without an explicit warning that this interface can expose tabs, session state, cookies, and arbitrary browser actions to local or nearby processes if misconfigured, users may enable a dangerous capability without understanding the risk.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
91% confidence
Finding

Falling back to npx --yes agent-browser without pinning an exact version makes execution depend on whatever package version the registry serves at runtime. In an agent skill context, this expands the trust boundary to the npm supply chain and can result in unexpected or malicious code execution if the package is compromised, typosquatted, or updated with breaking or unsafe behavior.

Content

No source excerpt is available for this finding.

subprocess module call

Medium
Category
Dangerous Code Execution
Confidence
70% confidence
Finding

subprocess module calls execute external commands. Without careful input validation, this enables command injection.

Content

Scanner excerpt · scripts/run_agent_browser_on_tabbit.py (reported line 100)May include surrounding context.

python
sys.stdout.write(shlex.join(command) + "\n")
        return 0

    completed = subprocess.run(command)
    return completed.returncode

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

Line L45 uses https://www.baidu.com as the example URL for browser operation. While only an example, it steers usage toward a specific locale/language context without explanation or user choice, which can conflict with language/locale policy expectations.

Content

No source excerpt is available for this finding.