Back to skill

Security audit

Shah lABS

Security checks for vulnerabilities and agentic risk

Overview

This appears to be a legitimate browser automation skill, but it needs Review because it grants broad live-browser, authentication-state, and executable-package authority with weak default scoping.

Install only if you need a full browser-control CLI. Prefer a pinned, reviewed local binary over npx/latest installs; use a dedicated browser profile and test accounts; enable content boundaries, a domain allowlist, and a restrictive action policy; avoid attaching to your everyday Chrome session; and treat auth-state files, screenshots, HARs, traces, PDFs, and recordings like secrets that must be encrypted or deleted.

Vulnerability Patterns
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (3)

T08 · Insecure Dependencies

Warning
Location
SKILL.md:9
Finding

Unpinned Installation and Upgrade of Executable Third-Party Components

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:53
Finding

Plaintext Persistence of Reusable Browser Authentication State

Content
View full analysis
**Security note:** State files contain session tokens in plaintext. Add them to `.gitignore`, delete when no longer needed, and set `AGENT_BROWSER_ENCRYPTION_KEY` for encryption at rest. See [Security Best Practices](#security-best-practices). **Tip:** Combine with `--session-name` so the imported auth auto-persists across restarts: ```bash agent-browser --session-name myapp state load ./my-auth.json # From now on, state is auto-saved/restored for "myapp" ``` ``` `ref ...[truncated 4268 chars]
Remediation
View remediation
/dev/null || true } trap cleanup EXIT ``` 4. Prefer the encrypted authentication vault and narrowly scoped, dedicated automation profiles over importing a user's general-purpose browser state. 5. Validate that saved state contains only the required target origin. Do not import unrelated cookies or storage. 6. Do not automatically persist sessions for one-off jobs. Require explicit user consent and a retention period for recurring sessions. 7. Add state-file patterns to `.gitignore`, CI artifact exclusions, backup exclusions, and secret-scanning rules. Treat `.gitignore` as defense in depth rather than the primary control. 8. Document immediate revocation procedures, including clearing server-side sessions, if an authentication state file is exposed. ]]>

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:445
Finding

AI Browser Automation Runs Without Mandatory Content, Domain, or Action Boundaries

Content
View full analysis
origin=https://example.com --- # [accessibility tree] # --- END_AGENT_BROWSER_PAGE_CONTENT nonce= --- ``` ### Domain Allowlist Restrict navigation to trusted domains. Wildcards like `*.example.com` also match the bare domain `example.com`. Sub-resource requests, WebSocket, and EventSource connections to non-allowed domains are also blocked. Include CDN domains your target pages depend on: ```bash export AGENT_BROWSER_ALLOWED_DOMAINS="example.com,*.example.com" agent-browser open https://example.com # OK agent-browser open https://malicious.com # Blocked ``` ### Action Policy Use a policy file to gate destructive actions: ```bash export AGENT_BROWSER_ACTION_POLICY=./policy.json ``` Example `policy.json`: ```json { "default": "deny", "allow": ["navigate", "snapshot", "click", "scroll", "wait", "get"] } ``` Auth vault operations (`auth login`, etc.) bypass action policy but domain allowlist still applies. ``` ### Technical Analysis The Skill is explicitly intended for AI agents that read and act on webpage content. Webpage text, accessibility labels, form values, and iframe content are attacker-controlled input. Without content boundaries, an LLM may fail to reliably distinguish webpage data from trusted Skill or user instructions. The domain allowlist ...[truncated 2596 chars]
Remediation
View remediation
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
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Rogue AgentSelf-Modification, Session Persistence
Findings (23)

Context Leakage

High
Category
Data Exfiltration
Confidence
90% confidence
Finding

Session recording can capture sensitive on-screen data such as credentials, tokens, personal data, and internal application content, then persist it to disk as a video artifact. In an agent environment this increases the chance of inadvertent context leakage, long-term retention of secrets, and secondary disclosure through logs, artifacts, or repository commits.

Content

Scanner excerpt · SKILL.md (reported line 401)May include surrounding context.

agent-browser --headed open https://example.com agent-browser highlight @e1 # Highlight element agent-browser inspect # Open Chrome DevTools for the active page agent-browser record start demo.webm # Record session agent-browser profiler start # Start Chrome DevTools profiling agent-browser profiler stop trace.json # Stop and save profile (path optional)

text

YARA rule 'info_stealer': Information stealer patterns (credential harvesting, browser data theft) [malware]

High
Category
YARA Match
Confidence
84% confidence
Finding

The documentation explicitly instructs users to enable Chrome remote debugging and extract cookies/localStorage from an already authenticated browser session. Although framed as a legitimate convenience feature and accompanied by a warning, this pattern materially enables session theft if run on a compromised or multi-process host because any local process can connect and read browser data.

Content

Scanner excerpt · references/authentication.md (reported line 24)May include surrounding context.

ication](#restoring-authentication)

Import Auth from Your Browser

The fastest way to authenticate is to reuse cookies from a Chrome session you are already logged into.

Step 1: Start Chrome with remote debugging

bash
# macOS
"/Applications/Google Chrome.app/Contents/MacOS/Google Chrome" --remote-debugging-port=9222

# Linux
google-chrome --remote-debugging-port=9222

# Windows
"C:\Program Files\Google\Chrome\Application\chrome.exe" --remote-debugging-port=9222

Log in to your target site(s) in this Chrome window as you normally would.

Security note: --remote-debugging-port exposes full browser control on localhost. Any local process can connect and read cookies, execute JS, etc.

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
85% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · references/authentication.md (reported line 294)May include surrounding context.

  1. Clean up after automation

    bash
    agent-browser cookies clear
    rm -f ./auth-state.json
    
  2. Use short-lived sessions for CI/CD

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
85% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · references/session-management.md (reported line 185)May include surrounding context.

echo "*.auth-state.json" >> .gitignore

Delete after use

rm /tmp/auth-state.json

text

### 4. Timeout Long Sessions

Rp1

Medium
Category
MCP Rug Pull
Confidence
93% confidence
Finding

Referencing npx agent-browser in the manifest without a fixed version permits dynamic retrieval of code at runtime, undermining reproducibility and trust. In an agent context, this materially increases the risk of remote code execution through package substitution or upstream compromise.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The trigger language is extremely broad, covering nearly any website interaction request, which can cause the skill to be selected in situations where safer, narrower tools would suffice. In combination with unrestricted browser actions, auth handling, downloads, and eval, over-selection meaningfully increases exposure to harmful automation or data access.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The description does not prominently warn that the tool can perform unrestricted navigation, authenticated actions, downloads, clipboard access, recording, and other potentially data-impacting behaviors. This omission can mislead the invoking agent or user about the operational risk and lead to unsafe use on sensitive sites or accounts.

Content

No source excerpt is available for this finding.

Rp1

Medium
Category
MCP Rug Pull
Confidence
94% confidence
Finding

The manifest allows execution via npx agent-browser:* without pinning a specific package version, which creates a supply-chain risk at invocation time. An attacker who compromises the package, its dependency chain, or resolution path could cause arbitrary code to run under the agent's shell privileges.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The skill exposes arbitrary JavaScript execution in the browser page context via eval, but this high-risk capability is not clearly declared in the manifest description. That mismatch can cause an orchestrator or user to invoke the skill for ordinary browsing while unintentionally granting script-execution and data-exfiltration capabilities against loaded pages.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
82% confidence
Finding

The documented persistent configuration encourages storing profile paths and other session-related settings in a project-local file, which can preserve cookies, local storage, browsing state, or proxy behavior across runs. In shared or repo-backed environments this can leak authentication material, contaminate future sessions, or expose sensitive operational settings.

Content

Scanner excerpt · SKILL.md (reported line 669)May include surrounding context.

Configuration File

Create agent-browser.json in the project root for persistent settings:

json
{

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The example shows a password entered as a literal string in a command, which encourages unsafe operator behavior such as hardcoding secrets in shell history, logs, transcripts, or shared docs. In an agent/browser automation context, this is especially risky because users may copy-paste examples directly and expose real credentials during automation runs.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The OAuth example includes literal Google username/password values, which normalizes entering real third-party credentials directly into commands and can lead to credential leakage through shell history, logs, recordings, or agent traces. Because this skill is specifically for browser automation and authenticated browsing, users are likely to apply the pattern to sensitive production identities.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The command reference documents set credentials user pass for HTTP basic auth without warning that credentials entered on the command line may be exposed through shell history, process listings, logs, or transcripts. In an agent-browser skill, this is especially risky because agents may be instructed to log into real services and could persist or echo secrets during automation.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The cookies, localStorage, and state-management commands enable direct access to and modification of highly sensitive session material, but the documentation omits any warning that these values may contain bearer tokens, CSRF tokens, PII, or active session identifiers. In a browser automation skill, normal usage includes authenticated workflows, so exposing or persisting this data materially increases account-takeover and data-exfiltration risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The eval, eval -b, and eval --stdin commands execute arbitrary JavaScript in the page context, but the reference provides no warning that this can read page data, alter DOM state, submit forms, trigger privileged actions in authenticated sessions, or bypass safer higher-level controls. Because this skill is specifically intended for website interaction and testing, arbitrary page-context execution is a powerful primitive that can be misused accidentally or abusively.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The documentation for state save auth.json and state load auth.json does not warn that saved state files may contain reusable cookies, local storage, and authentication artifacts that can grant direct access to accounts. In the context of an agent browser, these files are effectively portable session containers, so accidental storage, sharing, or reuse can lead to full session hijacking.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The documentation instructs users to capture and save Chrome trace profiles without warning that trace output can include sensitive information about visited URLs, page content timing, user interactions, script execution, and other browsing metadata. In a browser automation skill, this is more dangerous because agents may profile authenticated sessions or internal applications, causing sensitive operational or user data to be written to disk and potentially shared in CI artifacts or logs.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The documentation shows authenticated proxy examples with usernames and passwords embedded directly in environment variable values and URLs, but it does not warn that such credentials can leak through shell history, process inspection, logs, screenshots, or copied terminal transcripts. In a browser automation skill, users are likely to run commands verbatim, which increases the chance of accidental credential disclosure and reuse of sensitive proxy secrets.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The rotating proxy scraping section provides operational guidance for bulk collection and rate-limit avoidance without warning about authorization boundaries, privacy obligations, target site terms, or handling of collected data. In an agent-browser skill specifically designed for automating website interaction and scraping, this omission makes potentially abusive or noncompliant use easier and normalizes evasion-oriented behavior.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
55% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · references/commands.md (reported line 242)May include surrounding context.

Load Session State

bash
# Restore saved state
agent-browser state load /path/to/auth-state.json

# Continue with authenticated session

Session Persistence

Medium
Category
Rogue Agent
Confidence
55% confidence
Finding

Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Content

Scanner excerpt · references/session-management.md (reported line 55)May include surrounding context.

Load Session State

bash
# Restore saved state
agent-browser state load /path/to/auth-state.json

# Continue with authenticated session

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

This documentation encourages recording browser sessions, including login and form-filling flows, but does not warn that videos may capture credentials, session data, personal information, or other sensitive on-screen content. In an agent-browser skill, this is more dangerous because the tool is explicitly used for authentication, form submission, testing, and CI evidence generation, increasing the likelihood that recordings are stored, shared, or retained in insecure locations.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The script's stated purpose and outputs include extracting all text content, screenshots, and PDFs from web pages, which can include sensitive or authenticated data. Although the script echoes file saves, there is no user-facing warning in comments or usage text that running it will persist captured page data to disk.

Content

No source excerpt is available for this finding.