Back to skill

Security audit

Review

Security checks for vulnerabilities and agentic risk

Overview

The skill is a mostly coherent PR review helper, but it can directly edit workspace files during what is described as a review flow.

Install only if you are comfortable with a review skill that can edit files automatically. Before using it on important branches, require review-only mode or explicit approval before every fix, and inspect the resulting diff carefully.

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

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:112
Finding

Review Skill Performs Workspace Modifications Without Explicit User Approval

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:112-115 and SKILL.md:159-160
Vulnerability Type: Unauthorized modification beyond the declared review scope
Risk Level: Medium

Vulnerable Code

markdown
### 5b: Auto-fix all AUTO-FIX items
Apply each fix directly. Output one-line summary per fix:
`[AUTO-FIXED] [file:line] Problem → what you did`

The behavior is reiterated by the following instruction:

markdown
- Fix-first, not read-only. AUTO-FIX applied directly. ASK items need user approval.

Technical Analysis

The Skill is described as a pre-landing review mechanism that analyzes a branch diff. However, its operational instructions expand this review-only purpose by requiring the Agent to apply changes directly to the workspace.

User approval is required only for findings internally classified as ASK. Findings classified as AUTO-FIX are modified immediately. Because the Skill itself controls this classification, the distinction permits workspace changes without explicit user consent for each proposed patch.

This violates least-privilege and explicit-consent principles: a task that can be completed through read-only repository inspection should not implicitly exercise write access. Although the Skill does not obtain additional operating-system privileges, it uses available filesystem modification privileges beyond the expected scope of a code-review request.

Attack Path

  1. A user invokes the Skill with a review-only request, such as asking it to review a pull request.
  2. The Skill reads the branch diff and identifies an issue.
  3. The Skill internally classifies the issue as AUTO-FIX.
  4. Under the instructions in SKILL.md, the Agent edits the affected project file immediately.
  5. The workspace or current branch is changed without the user first reviewing or approving the patch.
  6. If the generated correction is faulty, the unauthorized edit may introduce regr ...[truncated 868 chars]
Remediation
View remediation

Remediation Suggestions

  1. Make the default review workflow strictly read-only.
  2. Replace the direct-application instruction with generation of a proposed patch or concise remediation description.
  3. Require explicit user approval before every filesystem modification, including findings classified as AUTO-FIX.
  4. Introduce a clearly documented opt-in mode, such as --fix, if automatic modifications are required.
  5. Before applying an approved fix, display the affected files, proposed changes, and expected behavioral impact.
  6. After modification, report every changed file and provide a diff so the user can verify or revert the patch.
  7. Update the Skill description to disclose any write behavior accurately if automatic fixing remains supported.
  8. Do not allow an internal severity or fixability classification to substitute for user authorization.
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (5)

Description-Behavior Mismatch

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill is presented as a PR review capability, but its instructions explicitly require directly modifying code via auto-fixes during the review flow. That creates a dangerous mismatch between user expectations and actual behavior: a user invoking a 'review' skill may unknowingly authorize writes to the repository, including changes beyond what they intended to inspect.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The description markets the skill as a review tool but omits that it will directly apply AUTO-FIX changes. Hidden write behavior undermines informed consent and can cause users to trigger code changes when they expected analysis only; in a security-sensitive repository, this can alter code, tests, or configuration without a clearly separate approval step.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

Broad trigger phrases like 'code review' or 'check my diff,' combined with proactive suggestion behavior, increase the likelihood that the skill activates in situations where the user did not intend this specific workflow. In this skill, that risk is amplified because the workflow can perform repository reads and even code modifications, making accidental invocation materially more dangerous.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
84% confidence
Finding

The stated purpose is analyzing the current branch diff for structural issues such as SQL safety and trust-boundary problems. The instructions additionally require reading TODOS.md, root documentation files, PR descriptions, and commit history to perform TODO cross-references and stale-documentation checks, which expands the skill into project-management and documentation-audit tasks not disclosed in the manifest.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The optional handoff to Codex introduces an external review path that is not clearly necessary for the stated purpose and may expose plan or repository context to another tool. Even though it asks the user first, embedding a tool-transfer workflow inside a review skill increases the chance of unintended data sharing and expands the trust boundary.

Content

No source excerpt is available for this finding.

Static analysis

No suspicious patterns detected.