khe-brain-harshit
Fail
Audited by Gen Agent Trust Hub on Nov 17, 2026
Risk Level: CRITICALEXTERNAL_DOWNLOADSCOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONOBFUSCATION
Full Analysis
- [EXTERNAL_DOWNLOADS]: The skill corpus (references/his-story.md and references/sources.md) contains a reference to whoburnedmore.com, which is flagged by automated scanners. In the context of the skill, this is a factual reference to a leaderboard for AI coding token usage and is not used for automated code fetching. The documentation also mentions the smashandclash package available via npx.
- [COMMAND_EXECUTION]: Instructions in SKILL.md direct the AI agent to execute a provided Node.js script (scripts/brain-check.mjs) during the rehearsal phase. This script is used locally to lint drafts for formatting, citation accuracy, and persona consistency.
- [INDIRECT_PROMPT_INJECTION]: The skill has an attack surface for indirect injection as it processes user queries into persona-based drafts. Evidence Chain: 1) Ingestion points: User requests are processed via the workflow defined in SKILL.md. 2) Boundary markers: The skill mandates specific output headers (Source:, Rests on:, Not in his posts:) to delimit information. 3) Capability inventory: The agent can execute a local Node.js script and read/write session files. 4) Sanitization: The validation script (scripts/brain-check.mjs) performs regex-based checks to filter out hedge words and ensure claims are backed by the provided references.
- [OBFUSCATION]: The internal validation script defines restricted terms using string concatenation (e.g., 'dis' + 'tilled'). This common practice prevents the utility from flagging its own source code during self-testing and does not represent a malicious obfuscation attempt.
Recommendations
- Contains 1 malicious URL(s) - DO NOT USE
Audit Metadata