Back to skill

Security audit

Fast Agent Browser

Security checks for vulnerabilities and agentic risk

Overview

This appears to be a real browser automation skill, but it needs Review because it can run page scripts, browse arbitrary URLs, upload files, and capture page data while overstating its safety controls.

Install only if you are comfortable giving the agent broad browser control. Avoid using it on authenticated, internal, financial, admin, or private sites unless you explicitly approve each action, and treat screenshots, extracted HTML/text, uploaded files, cookies, and auth state as sensitive. Prefer isolated test accounts and pinned dependencies before production use.

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)

T09 · Insecure Skill Coding Practices

Error
Location
src/browser.py:44
Finding

Unrestricted URL Navigation Enables Server-Side Request Forgery and Local Network Access

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
src/browser.py:194
Finding

Caller-Controlled Values Are Interpolated into Executable JavaScript

Content
View full analysis
Remediation
View remediation
document.body.innerText.includes(expected)", text, timeout=timeout, ) ``` For element scrolling, use a locator or element handle: ```python locator = self.page.locator(selector) locator.evaluate( "(element, amount) => element.scrollBy(0, amount)", pixels if direction == "down" else -pixels, ) ``` Additional hardening should include: 1. Enforce `pixels` as an integer and set a reasonable absolute range. 2. Restrict `direction` and `state` to explicit enumerated values. 3. Let Playwright parse selectors instead of embedding them into JavaScript. 4. Add tests containing quotes, backslashes, newlines, template-literal syntax, and JavaScript statement separators. 5. Clearly document the separate explicit `eval()` method as arbitrary page-context execution rather than claiming all commands cannot execute arbitrary code. ]]>

T08 · Insecure Dependencies

Warning
Location
requirements.txt:1
Finding

Unpinned Dependencies and Global Latest-Package Installation Create Supply-Chain Exposure

Content
View full analysis
=1.40.0 click>=8.1.0 pydantic>=2.5.0 rich>=13.7.0 ``` `CONTRIBUTING.md`: ```markdown ## Before Opening an Issue 1. Install the latest version ```bash npm install -g agent-browser@latest ``` 2. Test the command in your terminal to isolate the issue ``` ### Technical Analysis Every Python dependency uses an open-ended minimum version. A future installation may therefore retrieve package versions that did not exist when the Skill was reviewed. No lockfile or package hashes are provided to ensure that users install the audited artifacts. The contributor instructions additionally recommend installing `agent-browser@latest` globally from npm. That package is separate from the bundled Python implementation and can change at any time. Global npm installation may execute package lifecycle scripts and places files in a broadly accessible command location. This is a supply-chain weakness rather than evidence that the currently named packages are malicious. ### Attack Path 1. A user follows the installation or troubleshooting documentation. 2. The package manager resolves newer packages because versions are open-ended or explicitly use `@latest`. 3. A compromised account, malicious newly resolved package version, or compromised transitive dependency supplies altered package content. 4. Installation hooks or imported dependency code executes with the installing user's privileges. 5. The malicious dependency can access files, credentials, network resources, and other data available to that user. ### Impact Assessment Impact depends on the privileges used for package installation. Installation under a normal account can compromise th ...[truncated 266 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
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
  • Rogue AgentSelf-Modification, Session Persistence
Findings (27)

Self-Modification

High
Category
Rogue Agent
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.

Content

Scanner excerpt · CONTRIBUTING.md (reported line 60)May include surrounding context.

md
## 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

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

The class exposes a direct eval(javascript) wrapper over page.evaluate, allowing any caller of the skill to execute arbitrary JavaScript in the context of whatever page is open. In an agent setting, this bypasses the safer higher-level browser actions and can be abused to extract sensitive page data, manipulate forms invisibly, or trigger unsafe workflows on trusted sites.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The README explicitly documents saving/loading auth state and manipulating cookies, but does not warn that these artifacts can contain live session tokens and account credentials. In an AI-agent/browser-automation context, this increases the risk of unsafe persistence, accidental reuse, leakage to logs/workspaces, or cross-task account compromise if operators treat session files as harmless.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The examples demonstrate login, form filling, checkbox selection, and submit actions without warning that these commands can trigger real-world state changes on remote systems. For AI agents, this omission can normalize blind execution of side-effecting actions, leading to unintended submissions, account actions, purchases, or data modification when examples are adapted to production targets.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The overview and feature description present browser automation as generic and AI-friendly without warning that it can navigate real sites, submit forms, click destructive actions, or alter remote state. In an agent setting, omission of these guardrails increases the risk that autonomous workflows perform unintended transactions, account changes, or policy-violating actions on live services.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The examples normalize use with credentials, uploads, screenshots, and extraction workflows but provide no privacy, secrecy, or data-handling warnings. That is risky because agents may capture secrets, upload sensitive local files, or exfiltrate page contents/screenshots from authenticated sessions without clear constraints.

Content

No source excerpt is available for this finding.

Unrestricted Tool Access

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill grants unrestricted tool access without appropriate constraints. An agent with unfettered tool access can perform arbitrary actions including file modification, network requests, and code execution.

Content

Scanner excerpt · SKILL.md (reported line 338)May include surrounding context.

md
### Safe Commands

All commands are safe and do not execute arbitrary code:

- No shell injection possible
- No command injection possible

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The documentation makes a broad safety claim that all commands are safe and do not execute arbitrary code, yet the skill explicitly exposes an eval command that runs user-supplied JavaScript in the browser context. This can mislead downstream agents or users into treating the tool as harmless, increasing the chance of unsafe automation against sensitive pages, token-bearing sessions, or untrusted sites.

Content

No source excerpt is available for this finding.

Context Window Stuffing

Medium
Category
Memory Poisoning
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.

Content

Scanner excerpt · README.md (reported line 73)May include surrounding context.

md
print("  open <url>              Open a URL")
        print("  snapshot [-i] [-c] [-d] Get accessibility tree")
        print("  click <selector>        Click element")
        print("  fill <sel> <text>       Fill input field")
        print("  type <sel> <text>       Type text")
        print("  screenshot [path]       Take screenshot")
        print("  get_text <sel>          Get text content")

Context Window Stuffing

Medium
Category
Memory Poisoning
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.

Content

Scanner excerpt · SKILL.md (reported line 68)May include surrounding context.

md
print("  open <url>              Open a URL")
        print("  snapshot [-i] [-c] [-d] Get accessibility tree")
        print("  click <selector>        Click element")
        print("  fill <sel> <text>       Fill input field")
        print("  type <sel> <text>       Type text")
        print("  screenshot [path]       Take screenshot")
        print("  get_text <sel>          Get text content")

Context Window Stuffing

Medium
Category
Memory Poisoning
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.

Content

Scanner excerpt · agent_browser.py (reported line 20)May include surrounding context.

python
print("  open <url>              Open a URL")
        print("  snapshot [-i] [-c] [-d] Get accessibility tree")
        print("  click <selector>        Click element")
        print("  fill <sel> <text>       Fill input field")
        print("  type <sel> <text>       Type text")
        print("  screenshot [path]       Take screenshot")
        print("  get_text <sel>          Get text content")

Context Window Stuffing

Medium
Category
Memory Poisoning
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.

Content

Scanner excerpt · docs/api.md (reported line 215)May include surrounding context.

md
print("  open <url>              Open a URL")
        print("  snapshot [-i] [-c] [-d] Get accessibility tree")
        print("  click <selector>        Click element")
        print("  fill <sel> <text>       Fill input field")
        print("  type <sel> <text>       Type text")
        print("  screenshot [path]       Take screenshot")
        print("  get_text <sel>          Get text content")

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The CLI advertises an eval <js> command that executes JavaScript, which is a safety-sensitive operation because it can modify page state or trigger unintended actions. In this file there is no confirmation prompt or warning text describing the risks of executing arbitrary code.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The CLI help and command branch describe this action as 'type', which implies simulated keystrokes, but the implementation invokes agent.fill(args[0], args[1]) instead. This is an active contradiction between the documented command behavior and the code path executed.

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 · agent_browser.py (reported line 156)May include surrounding context.

python
elif command == 'install':
            import subprocess
            print("Installing Playwright browsers...")
            result = subprocess.run([sys.executable, '-m', 'playwright', 'install'], capture_output=True, text=True)
            if result.returncode == 0:
                print("✓ Playwright browsers installed successfully")
            else:

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The documentation encourages screenshots and page-content capture without warning that these operations may collect sensitive data such as credentials, tokens, personal information, or confidential on-screen content. In an agent/browser automation context, users may apply these examples to authenticated or internal pages, increasing the chance of unintended data exposure through saved files or logs.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
86% confidence
Finding

The upload method sends local files from user-supplied paths into a browser page via set_input_files, which can expose user data to a remote site. The code contains no confirmation prompt, logging, or warning comment/docstring explaining this privacy-impacting action beyond the neutral method name.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

Arbitrary JavaScript execution is exposed without any warning, gating, or policy checks. In a browser agent, undisclosed script execution increases the risk of hidden DOM scraping, token extraction from page context, and actions that are not apparent from the higher-level API surface.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
77% confidence
Finding

The screenshot [path] command writes an image file to disk, which is a file-modifying operation. Although success is printed after execution, the help/description does not warn beforehand that invoking the command will create or overwrite a local file.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
80% confidence
Finding

The install command launches a subprocess to install Playwright browsers, which changes the local system state. While the code prints a message at runtime, the command listing does not warn users that invoking install will download and install browser binaries.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
94% confidence
Finding

The 'Persistent Sessions' section shows a flow where a second BrowserAgent instance is 'Already authenticated', but the same example states 'Save state (cookies, localStorage)' and marks it as '(Implementation needed)'. This actively misleads readers about the skill's actual behavior by implying persisted authentication/session reuse that the documentation itself indicates is not implemented.

Content

No source excerpt is available for this finding.

Unpinned Dependencies

Low
Category
Supply Chain
Confidence
96% confidence
Finding

The dependency is specified with a lower-bound constraint only, which makes builds non-reproducible and allows later releases to be installed without review. This increases supply-chain risk because a vulnerable or breaking upstream version could be pulled in at install time.

Content

Scanner excerpt · requirements.txt (reported line 1)May include surrounding context.

text
playwright>=1.40.0
click>=8.1.0
pydantic>=2.5.0
rich>=13.7.0

Unpinned Dependencies

Low
Category
Supply Chain
Confidence
98% confidence
Finding

The click dependency is unpinned, so installations may resolve to different versions over time, including versions with known security defects. Because command-line frameworks can influence command execution paths, leaving this dependency floating meaningfully increases supply-chain and exposure risk.

Content

Scanner excerpt · requirements.txt (reported line 2)May include surrounding context.

text
playwright>=1.40.0
click>=8.1.0
pydantic>=2.5.0
rich>=13.7.0

Unverifiable Dependency: click has 1 known advisory(ies) (CVE-2026-7246 (Pallets Click, versions 8.3.2 and below, contain a command injection vulnerabili)), but the manifest does not pin a version, so it is unknown whether the installed release is affected

Low
Category
Supply Chain
Confidence
95% confidence
Finding

The manifest does not pin click, and there is at least one known advisory affecting some versions, so it is impossible to verify from this file whether deployment will resolve to a safe release. In this context, that uncertainty is itself dangerous because a CLI-oriented package with command injection history could expose users if an affected version is installed.

Content

No source excerpt is available for this finding.

Unpinned Dependencies

Low
Category
Supply Chain
Confidence
97% confidence
Finding

The pydantic dependency is unpinned, which allows future installs to select different versions, potentially including releases affected by denial-of-service or parsing-related vulnerabilities. For a validation library that processes untrusted input, version drift can directly affect security posture.

Content

Scanner excerpt · requirements.txt (reported line 3)May include surrounding context.

text
playwright>=1.40.0
click>=8.1.0
pydantic>=2.5.0
rich>=13.7.0