The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's primary function involves processing untrusted codebase artifacts including source code, configuration files, and dependency manifests (SKILL.md). This ingestion of external data presents a surface for indirect prompt injection. The skill mitigates this through defined boundary markers, specifically a mandatory confidence gate and an adversarial revalidation process (references/verification.md). Available capabilities are restricted through an approval-gated active phase for tool execution (SKILL.md Step 2). The skill also provides explicit instructions for sanitizing control characters and normalizing input to prevent exploitation of these sinks (§1 in checklists.md).
[EXTERNAL_DOWNLOADS]: The skill provides instructions for utilizing external security scanners (e.g., Semgrep, CodeQL) and third-party rulesets from reputable organizations like Trail of Bits. All network-based operations, including registry resolution and tool execution, are explicitly gated behind a mandatory user approval process described in the skill's work process (Step 2).
[OBFUSCATION]: The audit checklist (references/checklists.md) contains zero-width characters (U+200B, U+200D, U+FEFF) and bidirectional control characters (U+2066, U+2069, U+202E). These characters are utilized solely as literal example patterns within the documentation to teach the agent how to identify steganographic prompt injection and control-character smuggling in audited code; they do not obfuscate instructions or logic within the skill itself.
[SAFE]: The skill adheres to security best practices by implementing least-privilege principles, requiring explicit authorization for all active probes, and providing clear remediation guidance without attempting to perform automated, unreviewed modifications to the user's environment.

awesome-security-audit