The Agent Skills Directory

[Indirect Prompt Injection Surface]: The docker agent eval functionality ingests data from JSON session files to simulate agent interactions. If a user incorporates evaluation files from untrusted sources, there is a potential surface for indirect instructions to be processed by the agent during the test run. The skill encourages generating these files through the agent's own interactive interface, which helps maintain control over the data provenance.
[Dynamic Execution]: The evaluation framework allows for a setup shell script to be defined within the session JSON, which is executed inside a container to prepare the environment. While this is an intended feature for environment preparation, users should exercise standard caution when using evaluation suites provided by external parties to ensure the setup scripts align with expected behavior.

docker-agent-deploy