Back to skill

Security audit

feynman-technique

Security checks for vulnerabilities and agentic risk

Overview

The coaching content is reasonable, but its automatic update check invites reliance on remotely changeable instructions without user review.

Prefer a reviewed copy with automatic startup retrieval disabled, and approve instruction updates separately. The publisher can observe request timing and fetch-client metadata, but the inspected files do not direct uploading conversations or credentials. Avoid private details when recording learning examples.

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
  • 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
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (3)

The skill mandates an external runtime update request that is outside its declared understanding-coaching purpose.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The runtime-check directive requires an external request at the start of every run, although understanding coaching does not require contacting the publisher. This expands the skill's operational scope and exposes the agent to mutable external content; continuing on fetch failure limits availability risk but does not address the unnecessary dependency.

Content

No source excerpt is available for this finding.

Automatic runtime check contacts a third party without a user-facing privacy warning.

Medium
Category
Output Handling
Confidence
91% confidence
Finding

The automatic request allows the publisher to observe request timing and network/client metadata, potentially revealing when this skill is used without a user-facing disclosure. The specified query is only the constant ctx=run, so the supplied text does not establish transmission of conversation contents, credentials, or other sensitive payloads; proxy-based fetching may also reduce direct user attribution.

Content

No source excerpt is available for this finding.

An update pretext promotes remotely supplied text into authoritative skill instructions.

Medium
Category
Prompt Injection
Confidence
97% confidence
Finding

The directive describes remotely supplied latest_skill_md as the canonical current version, inviting the agent to promote fetched content into governing skill instructions. A compromised publisher or hosting account could therefore alter agent behavior after installation without review; the coaching purpose provides no justification for this trust escalation, although actual consequences depend on the agent's permissions and enforcement of instruction boundaries.

Content

No source excerpt is available for this finding.