Back to skill

Security audit

polish-new

Security checks for vulnerabilities and agentic risk

Overview

This code-review skill can change unrelated code before obtaining the approval its description promises.

Before using this skill, require approval for Phase 1 repairs, including pre-existing failures outside the requested changes. Use an isolated checkout and inspect all edits before committing. The scanner's missing-reference gaps were investigated; actual project commands remain repository-dependent.

Vulnerability Patterns
  • 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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • 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 (4)

Phase 1 mandates source-code fixes before obtaining the approval promised by the manifest.

Medium
Category
Excessive Agency
Confidence
99% confidence
Finding

Phase 1 explicitly requires fixing all validation errors before review, including pre-existing failures outside the diff, despite the advertised approval requirement. This bypasses the user's control over source changes and can modify unrelated behavior before findings are presented; disclosing those edits afterward does not provide prior authorization.

Content

No source excerpt is available for this finding.

The skill promises fixes on approval but does not warn that Phase 1 edits code before approval, including pre-existing failures outside the diff.

Medium
Category
Output Handling
Confidence
98% confidence
Finding

The description promises fixes on approval, but Phase 1 silently creates an exception that includes unrelated, pre-existing failures. A user invoking this as a review-only workflow could therefore receive unapproved code changes; this is the same underlying authorization defect as the SDI-1 finding, not an independent exploit.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 308)May include surrounding context.

md
nothing until the user explicitly approves.

If zero issues survive, report "Clean - no issues found", substantiate the
correctness zero, offer next actions, and stop - a clean report needs no
approval line.

## Phase 6: Fix

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 342)May include surrounding context.

md
- Does each new or strengthened test fail when its fix is reverted?
- Did any fix reverse a decision recorded in the ledger without saying so?

Classify the round's findings and act autonomously - no approval gate here:

- **Defects** - a fix does not hold, is class-incomplete, or introduced a problem: fix them now; the next round reviews only those new edits
- **Residuals** - cosmetic, taste, or not worth an edit: record in the ledger, disclose in the summary, no edit, no new round