Guardian Angel

• Analyzed: May 12, 2026, 3:22 AM
• Findings: 1
• Version: 3.1.2
• Hash: 4acc28c5fd060f0bfce0254dc5d4c90c70d799aba34c73c42c9d6467fa447b15
• Source: palm
• External report: View on VirusTotal

Findings (1)

The 'Guardian Angel' skill is explicitly designed as a defensive security and ethics enforcement mechanism for AI agents, utilizing Thomistic virtue ethics. Its documentation (`SKILL.md`, `references/prompt-injection-defense.md`) employs advanced prompt engineering techniques (identity anchoring, coherence sensing) to make the AI agent resilient to prompt injection attacks. The plugin code (`plugin/`) implements a `before_tool_call` hook to block intrinsically evil actions (e.g., `rm -rf /`, fork bombs, harmful image generation) and escalates high-stakes actions for human approval using a nonce-based system. While the overall intent is clearly defensive, the nonce generation in `plugin/src/store.ts` uses `randomBytes(4).toString("hex")`, providing only 32 bits of entropy. This low entropy, though used for short-lived, one-time, parameter-bound approvals, represents a cryptographic vulnerability that could potentially be exploited by a sophisticated attacker within the approval window, making the skill 'suspicious' rather than 'benign'.