The Agent Skills Directory

[DYNAMIC_EXECUTION]: The example scripts examples/anthropic-examples.py, examples/openai-examples.py, and examples/vercel-ai-examples.ts utilize the eval() function to process mathematical expressions generated by the LLM. While these files are intended as simulations and include explicit comments advising against production use, they demonstrate a pattern where untrusted input provided by the LLM is executed as code. This creates a potential vector for remote code execution (RCE) if the examples are run in an unsandboxed environment.
[INDIRECT_PROMPT_INJECTION]: The skill documents patterns for interpolating user input directly into prompts (e.g., in SKILL.md and references/few-shot-learning.md). This creates a vulnerability surface for indirect prompt injection. The skill includes relevant educational sections on input sanitization, the use of boundary markers (like XML tags), and instructions to ignore embedded commands to help mitigate these risks.
[PROMPT_INJECTION]: Several files, including SKILL.md and scripts/prompt-validator.py (referenced), contain common prompt injection strings such as 'ignore previous instructions'. A manual review confirms these are used in the context of identifying 'dangerous' patterns for sanitization or as negative examples in tutorials. These instances are benign and serve the skill's educational purpose rather than attempting to override the agent's behavior.

prompt-engineering