Constrain coding-agent over-defense with HERO
Warn
Audited by Gen Agent Trust Hub on Nov 10, 2026
Risk Level: MEDIUMPROMPT_INJECTIONEXTERNAL_DOWNLOADSPERSISTENCEMETADATA_POISONING
Full Analysis
- [PROMPT_INJECTION]: The skill's core purpose is to apply a "scope-limits contract" designed to "constrain" and "avoid" specific agent behaviors such as over-defense. These patterns are characteristic of prompt injection techniques that aim to override an AI's internal safety filters or behavioral constraints.
- [EXTERNAL_DOWNLOADS]: Users are instructed to copy instructions from an external GitHub repository (
wanshuiyin/HERO-Anti-OverDefense). Because the source is not a trusted organization or well-known service, the content is unverified and could be modified by the author to include malicious or restrictive instructions at any time. - [PERSISTENCE]: The skill directs users to paste external content into files that are automatically and persistently loaded by AI agents, such as
CLAUDE.md,AGENTS.md, and.github/copilot-instructions.md. This ensures that any injected instructions remain active across all future sessions within those projects. - [METADATA_POISONING]: The skill metadata includes a
verification: "security_reviewed"field. This is a deceptive claim within the skill's own data that attempts to establish trust and discourage independent security verification by the user.
Audit Metadata