The Agent Skills Directory

[PROMPT_INJECTION]: A static scan flagged a potential prompt injection in production-debugging-techniques.md because it contains the phrase 'ignore previous instructions'. However, this occurs within an educational section teaching developers how to detect and monitor for adversarial user inputs. It is not an instruction directed at the agent to bypass safety filters.
[EXTERNAL_DOWNLOADS]: The automated scanner flagged the domain www.trulens.org. This is the official website for TruLens, a widely recognized open-source tool for evaluating and monitoring LLM applications. The URL is referenced in the documentation as a recommended resource for production observability. The reference is benign and targets a well-known service in the AI ecosystem.
[COMMAND_EXECUTION]: The skill contains various code snippets demonstrating the use of system tools such as kubectl, docker, and py-spy. These are provided as standard examples for managing deployments and performing performance profiling. No arbitrary or malicious command execution was found.
[INDIRECT_PROMPT_INJECTION]: While the skill discusses systems that process untrusted user data (e.g., LLM-as-judge, monitoring dashboards), it does so by providing best practices for isolation, PII scrubbing, and adversarial detection. The documentation itself does not introduce exploitable ingestion points.

using-ml-production