The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and analyze data from external sources, such as AI model API responses and suspected model copy files.\n
Ingestion points: API query outputs, logits, labels, and local model weight files are identified as primary inputs in SKILL.md and decision-tree.md.\n
Boundary markers: The instructions do not define specific delimiters or instructions to disregard embedded commands when the agent processes or summarizes these external outputs.\n
Capability inventory: The skill environment includes Python execution (via torch and tensorflow), network requests (via requests), and file system operations (for checkpointing and report generation).\n
Sanitization: The skill lacks documentation regarding the validation or escaping of data retrieved from external APIs before it is incorporated into the agent's reasoning context.\n- [EXTERNAL_DOWNLOADS]: The skill requires the installation of several external Python libraries for its assessment tasks.\n
The instructions direct the user to install torch, tensorflow, adversarial-robustness-toolbox (ART), requests, and matplotlib using standard package managers.\n
These are well-known, industry-standard libraries used for machine learning and security research.\n- [COMMAND_EXECUTION]: The skill involves the execution of Python commands to perform technical evaluations.\n
Instructions include using python3 to verify installations and run assessment scripts for model distillation and adversarial sample generation.\n
This execution is limited to the defined scope of security forensics and model analysis.\n- [DYNAMIC_EXECUTION]: The skill methodology involves generating and running scripts at runtime to assess model behavior.\n
This includes training student models via distillation and calculating adversarial perturbations (FGSM/PGD) using the adversarial-robustness-toolbox.\n
The scripts are generated based on established security assessment frameworks.

re-ai-attack