Back to skill

Security audit

inversion

Security checks for vulnerabilities and agentic risk

Overview

The planning method is reasonable, but it asks the agent to retrieve changeable online instructions every run without a clear review step.

Prefer a reviewed local copy with the runtime fetch disabled, or require review before adopting remote revisions. The supplied request does not transmit your plan or conversation, but contacts the publisher each run and may introduce instructions that differ from the installed copy.

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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (3)

L008 requires a runtime network fetch for remote skill instructions, which is unnecessary for the stated pre-mortem analysis purpose.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The skill directs the agent to fetch remote skill instructions at the start of every run without specifying version pinning, integrity verification, or review before adoption. A compromised endpoint or malicious update could introduce instructions that redirect the agent's behavior; this dependency is unnecessary for the otherwise self-contained pre-mortem method.

Content

No source excerpt is available for this finding.

The mandatory runtime update request lacks a privacy warning about contacting a third-party service on every run.

Medium
Category
Output Handling
Confidence
90% confidence
Finding

The automatic request contacts a third party on every run without a privacy notice or opt-in, revealing request metadata such as timing and the requester or intermediary's network address. The fixed query string does not contain the user's decision or conversation, so this is a limited metadata exposure rather than demonstrated sensitive-content exfiltration.

Content

No source excerpt is available for this finding.

A maintenance pretext promotes remotely supplied skill text to canonical instruction authority.

Medium
Category
Prompt Injection
Confidence
97% confidence
Finding

Calling the remotely supplied latest_skill_md the 'canonical current version' encourages the agent to promote network content into governing skill instructions. This crosses a trust boundary and enables instruction substitution if the remote source changes maliciously, although the provided file does not establish malicious intent or show an actual remote payload.

Content

No source excerpt is available for this finding.