The Agent Skills Directory

[DYNAMIC_EXECUTION]: The file templates/tool-use-advanced.ts contains a mock calculatorTool that utilizes the JavaScript eval() function to evaluate mathematical expressions derived from AI input. While the code includes a clear warning that eval is dangerous and should not be used in production, the inclusion of executable templates using this method represents a security risk if implemented by developers without modification.
[INDIRECT_PROMPT_INJECTION]: The skill provides templates such as templates/nextjs-api-route.ts and templates/cloudflare-worker.ts that ingest external user messages and forward them to the Claude API. These templates lack instruction boundary markers or sanitization strategies, creating a surface for indirect prompt injection attacks where malicious data could influence intended application logic.
Ingestion points: Request bodies in templates/nextjs-api-route.ts and templates/cloudflare-worker.ts accept arbitrary message arrays from the client.
Boundary markers: None implemented in the templates to differentiate between developer instructions and external data.
Capability inventory: The templates provide network access to the Anthropic API and demonstrate dynamic code execution via the eval function in the tool example.
Sanitization: The skill documents issues with specific Unicode characters in SKILL.md but does not implement comprehensive input validation or filtering in the provided templates.
[OBFUSCATION]: SKILL.md (Issue #16) provides technical guidance for sanitizing Unicode line separators (U+2028 and U+2029) to prevent JSON parsing errors in tool results. This is a legitimate documentation of an API edge case rather than a malicious attempt to hide code.

claude-api