The Agent Skills Directory

[COMMAND_EXECUTION]: The script scripts/inference_optimizer.py uses the torch.load() function to analyze and benchmark user-provided PyTorch models. This function is vulnerable to unsafe deserialization because it uses the pickle module internally, which can execute arbitrary code during the loading process. An attacker could exploit this by providing a malicious model file designed to execute commands when analyzed by the skill.
[DATA_EXFILTRATION]: The scripts/dataset_pipeline_builder.py script uses the standard xml.etree.ElementTree library for parsing Pascal VOC XML annotations. This library is susceptible to XML External Entity (XXE) attacks. A maliciously crafted XML file in a dataset could be used to read sensitive local files from the environment or conduct server-side request forgery (SSRF) when the dataset is processed or analyzed.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection due to the way it ingests and processes untrusted data from external datasets (COCO JSON, YOLO TXT, and Pascal VOC XML).
Ingestion points: External data is read from local files by scripts/dataset_pipeline_builder.py and scripts/vision_model_trainer.py to calculate statistics and generate configurations.
Boundary markers: There are no boundary markers or instructions to the agent to ignore natural language commands embedded within the metadata or labels of the processed datasets.
Capability inventory: The skill possesses significant capabilities, including file system access and the ability to execute shell commands for training and optimization, making it a high-value target for injection attacks.
Sanitization: No sanitization or validation is performed on the string content of the processed files to prevent the inclusion of malicious instructions intended for the agent.

senior-computer-vision