The Agent Skills Directory

[COMMAND_EXECUTION]: The skill includes a mathematical calculator tool in scripts/basic_agent_with_tools.py that utilizes the eval() function to process input strings. Although a character allowlist (0123456789+-*/(). ) is implemented to restrict the execution environment, the use of eval() on data derived from LLM output remains a significant security surface for potential code execution.\n- [EXTERNAL_DOWNLOADS]: The Agentspan runtime is designed to fetch external binary components from remote infrastructure. As documented in SKILL.md and references/cli-quick-reference.md, the agentspan server start command downloads a JAR server file, and the NPM-based CLI fetches a Go binary from an Amazon S3 bucket upon initialization.\n- [PROMPT_INJECTION]: The skill documentation and examples include logic for detecting and mitigating prompt injection attacks. Specifically, references/patterns.md and scripts/guardrail_demo.py define guardrail functions that scan for malicious strings like "ignore previous instructions", "act as", and "system prompt". These are defensive implementations provided as part of the orchestration runtime and do not represent injection attempts by the skill itself.

agentspan