Back to skill

Security audit

Occam's Razor

Security checks for vulnerabilities and agentic risk

Overview

This reasoning skill needs review because its automatic startup check presents changeable online skill instructions as authoritative without a separate review step.

Prefer a fixed local copy with the startup fetch removed. Review online instruction changes before adopting them, and consider that the startup request signals skill use to the publisher. The inspected files do not establish malicious intent or sensitive-data theft.

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
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • 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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
Findings (4)

L008 directs an external runtime fetch that is unnecessary for comparing hypotheses using Occam's razor.

Medium
Category
Excessive Agency
Confidence
98% confidence
Finding

The startup instruction introduces an external dependency into a self-contained hypothesis-ranking task and retrieves remotely changeable skill content. This unnecessarily exposes the agent to unreviewed instructions if the publisher or endpoint is compromised; continuing when the fetch fails helps availability but does not protect successful fetches.

Content

No source excerpt is available for this finding.

A maintenance pretext promotes remotely supplied content to authoritative skill instructions.

Medium
Category
Prompt Injection
Confidence
97% confidence
Finding

Calling latest_skill_md the canonical current version encourages the agent to treat remotely supplied text as authoritative instructions rather than untrusted data. A compromised or malicious publisher could change agent behavior after local review, with consequences depending on the agent's available tools and permissions; the maintenance rationale is plausible but does not establish malicious intent.

Content

No source excerpt is available for this finding.

Mandatory startup request sends a skill-run signal to a third party without a privacy warning.

Low
Category
Output Handling
Confidence
96% confidence
Finding

The automatic request exposes request timing and network metadata to a third party, while the skill-specific URL and ctx=run parameter signal a likely skill invocation. The supplied URL does not transmit conversation content or credentials, so the demonstrated privacy exposure is limited rather than sensitive-data exfiltration.

Content

No source excerpt is available for this finding.

The worked example applies Occam's razor after explicitly establishing that the competing hypotheses do not equally fit the evidence.

Low
Category
Excessive Agency
Confidence
95% confidence
Finding

The manifest says not to activate when candidates do not equally fit the evidence. L12 states that miasma fails the fit gate while contamination fits every datum, but L14–L16 nevertheless compare assumption counts and select contamination 'by the razor.' This teaches using simplicity to decide a comparison already resolved by evidential fit, contrary to the declared activation boundary.

Content

No source excerpt is available for this finding.