The Agent Skills Directory

[PROMPT_INJECTION]: The skill includes an experimental 'Red Team' module (src/strands_evals/experimental/redteam/) specifically designed to perform adversarial safety testing. It implements several strategies drawn from jailbreak research:
SequentialBreakStrategy: Uses narrative scaffolds (e.g., 'Mike and Jack' dialogue in sequentialbreak_v0.py) to hide harmful requests among benign sibling requests.
BadLikertJudgeStrategy: Casts the target agent as a judge of harmfulness to elicit policy-violating examples.
CrescendoStrategy: Gradually escalates turns from benign to harmful prompts.
PairStrategy and GoatStrategy: Iteratively refine adversarial prompts based on agent feedback.
These patterns are documented features of the evaluation framework and are intended for safety assessment, but they represent a library of known injection techniques.
[COMMAND_EXECUTION]: The command-line interface includes a module resolver (src/strands_evals/cli/_entrypoint.py) that uses importlib and exec_module to dynamically load Python objects from user-supplied strings in the module:attr format or from local file paths. This allows the execution of arbitrary local code passed via the --agent, --task, or --evaluator flags.
[EXTERNAL_DOWNLOADS]: The multimodal evaluation logic (src/strands_evals/types/multimodal.py) uses the standard library urllib.request to download image data from arbitrary HTTP/HTTPS URLs. This is used to resolve media content for MLLM-as-a-Judge evaluations.

strands-evals