Back to skill

Security audit

Auto Study

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed browser automation tool for study platforms, but it has Review-level concerns because it can answer and submit academic work, retain screenshots and scores locally, reuse authenticated browser profiles, and includes an unsafe WSL sandbox-bypass fallback.

Install only if you are comfortable giving an agent control over an authenticated Chrome profile for supported learning sites. Use it only where automated answering and submission are allowed, avoid formal exams or graded work unless explicitly permitted, review before any submit action, and periodically delete stored screenshots, answer records, scores, and browser profiles. Do not rely on the WSL sandbox-bypass fallback; keep artifact writes inside the intended workspace and verify external browser dependencies before use.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (2)

T05 · Unauthorized Access and Privilege Escalation

Error
Location
references/runtime-wsl.md:18
Finding

Instruction to Bypass Sandbox Isolation During Artifact Writes

Content
View full analysis

Vulnerability Details

File Location: references/runtime-wsl.md:18
Vulnerability Type: Sandbox-boundary bypass
Risk Level: High

Vulnerable snippet:

markdown
- If a socket error occurs while writing task artifacts, try using a shell loop outside the sandbox.

Technical Analysis

The runtime guidance explicitly instructs the agent to move execution outside the sandbox if a socket error occurs while writing artifacts. A sandbox is a security boundary intended to constrain filesystem, process, network, and host access. Treating a routine write failure as authorization to leave that boundary violates least-privilege principles.

The instruction does not constrain the external shell loop, specify permitted paths, require user confirmation, or define an allowlist of operations. Consequently, the command ultimately generated by an agent could operate with the broader permissions of the host-side agent process. The exact privileges available depend on the surrounding runtime and user account.

Attack Path

  1. A task invokes the WSL-specific runtime guidance.
  2. An artifact write encounters, or appears to encounter, a socket-related error.
  3. The agent follows the documented fallback and constructs a shell loop outside the sandbox.
  4. The loop executes beyond the original isolation boundary.
  5. If generated paths or shell arguments are unsafe or attacker-influenced, the operation may access or modify host resources that were unavailable inside the sandbox.

Impact Assessment

Successful exploitation could expose host files, processes, or network resources available to the agent's operating-system account. It could also allow artifact-writing operations to modify data outside the intended workspace. The maximum scope is bounded by the privileges of the external shell process, but those privileges are broader than the sandboxed task legitimately requires.

Remediation
View remediation

Remediation Suggestions

  • Remove the instruction to execute a shell loop outside the sandbox.
  • Keep all artifact creation and retry operations within the configured sandbox and workspace.
  • Treat socket failures as recoverable errors and report them to the user if sandboxed retries fail.
  • Canonicalize and validate every artifact destination before writing it.
  • Restrict destinations to an explicit workspace allowlist and reject path traversal, symbolic-link escapes, UNC paths, and unexpected absolute paths.
  • If an external operation is operationally unavoidable, require explicit user approval and use a narrowly scoped helper with fixed arguments rather than generated shell code.
  • Log the approved operation and verify that the destination remains within the authorized artifact directory.

T08 · Insecure Dependencies

Warning
Location
SKILL.md:90
Finding

Unpinned External Agent Skill Dependency

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:90-93
Additional Location: README.md:76-79
Vulnerability Type: Mutable third-party Agent Skill dependency
Risk Level: Medium

Vulnerable snippet:

markdown
## Prerequisites

- Google Chrome (on Windows or macOS)
- [Agent Browser CLI](https://github.com/vercel-labs/agent-browser)
- [Agent Browser Skill](https://clawhub.ai/MaTriXy/agent-browser-clawdbot)

The corresponding installation documentation also states:

markdown
#### Install dependencies

- Google Chrome (Windows or macOS)
- [Agent Browser CLI](https://github.com/vercel-labs/agent-browser)
- [Agent Browser Skill](https://clawhub.ai/MaTriXy/agent-browser-clawdbot)

Technical Analysis

The project requires a separately maintained Agent Skill through a mutable external URL, but it does not identify an immutable version, release, commit, checksum, signature, or verified artifact. Agent Skills can influence tool usage and browser behavior, so this dependency has capabilities relevant to authenticated Chrome sessions.

Because the dependency is not pinned, the component installed by a future user can differ from the component that was reviewed when this project was audited. Compromise of the hosting account, registry entry, or upstream dependency could therefore alter the effective behavior of an otherwise unchanged installation.

No automatic retrieval or execution implementation is present in this repository; exploitation depends on a user or agent following the documented dependency-installation process.

Attack Path

  1. An attacker compromises or maliciously updates the referenced external Agent Skill or its distribution account.
  2. The content available through the mutable dependency URL is replaced with an altered version.
  3. A user follows this project's prerequisites and installs the current external version without immutable-version or integrity verificat ...[truncated 661 chars]
Remediation
View remediation

Remediation Suggestions

  • Pin the external Agent Skill to an immutable, audited release or commit.
  • Publish the expected cryptographic hash for the exact dependency artifact.
  • Require checksum or signature verification before installation.
  • Record the reviewed dependency version in SKILL.md and the installation documentation.
  • Define the minimum permissions required by the dependency and deny unrelated filesystem, command, and network access.
  • Maintain a dependency update process that requires review before changing the pinned version or hash.
  • Prefer a trusted registry that provides immutable releases, provenance attestations, and publisher verification.
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (11)

Ssd 4

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

Explicitly telling the agent not to infer quiz, exam, practice, or homework status from page labels normalizes ignoring obvious signals of assessment context. In a skill designed to answer questions and perform page actions on education platforms, this materially increases the risk of unauthorized academic task completion and deliberate circumvention of safety boundaries.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The file is a Chinese-language README variant and does provide an English alternative link, but the skill examples and instructions in this file assume Chinese interaction without any explicit user opt-in or stated locale limitation. Under the stated policy, language constraints should either be optional for the user or clearly justified as region-specific.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill description is broad enough to match many ordinary study-related requests, including tasks on browser-based learning platforms that may involve graded coursework. In context, this increases the chance the agent is invoked for sensitive academic actions without clear boundary checks, which can enable unauthorized assistance or policy bypass.

Content

No source excerpt is available for this finding.

Ssd 4

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The instruction to treat all pages as ordinary practice by default weakens safeguards around academic integrity-sensitive contexts. Because many real assessments are not explicitly declared by users, this default can steer the agent into completing graded work or restricted tasks without proper verification.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The skill instructs reuse of a persistent Chrome profile and creation of local study artifacts, but the top-level description does not clearly warn users that login state, screenshots, and answer records may be retained locally. This can expose session data, educational records, and potentially sensitive page content on shared or insufficiently protected systems.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The skill explicitly directs the agent to take a full-page screenshot, extract questions, and persist both the image and question/answer content to local workspace files. This creates unnecessary local retention of potentially sensitive educational content, user identifiers, grades, or copyrighted material without any minimization, consent, retention, or access-control guidance. In this context, the file-writing is operationally useful, but still risky because it expands the data exposure surface beyond the browser session.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill explicitly instructs the agent to capture full-page screenshots of every question and transcribe all exam content into markdown, creating a persistent local copy of assessment materials without any warning, consent check, or data-handling limits. In this context, the content likely includes copyrighted test items and potentially user-associated academic data, so silent capture and storage increases privacy, compliance, and misuse risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill directs the agent to apply answers and submit the assignment/exam, including handling a confirmation popup, but does not require an explicit final user approval before taking this irreversible action. Because submission can finalize answers, affect grades, and cannot always be undone, automating it without a clear checkpoint materially increases the risk of unintended academic and account-impacting actions.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
80% confidence
Finding

The natural-language content explicitly exposes only English and Simplified Chinese variants. Under the policy rule, forcing or effectively limiting language/locale without an explicit opt-in or documented justification can be a language policy concern.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
85% confidence
Finding

The workflow instructs the agent to create a task directory and record all questions in a markdown file, which affects user data/storage. The skill description does not include any warning that it will create local files or persist exam content.

Content

No source excerpt is available for this finding.

Missing User Warnings

Low
Category
Not specified by scanner
Confidence
89% confidence
Finding

These steps instruct the agent to save a screenshot of the result page and append the final score to markdown, which stores potentially sensitive academic data. The markdown does not warn users that result images and scores will be persisted.

Content

No source excerpt is available for this finding.

Static analysis

No suspicious patterns detected.