senior-computer-vision

Warn

Audited by Gen Agent Trust Hub on Apr 2, 2026

Risk Level: MEDIUMREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONPROMPT_INJECTION
Full Analysis
  • [REMOTE_CODE_EXECUTION]: The script inference_optimizer.py utilizes torch.load() to process model files. This function uses the pickle module internally, which can be exploited to execute arbitrary code if a maliciously crafted model file is loaded by the user.\n- [DATA_EXFILTRATION]: The dataset_pipeline_builder.py script parses Pascal VOC annotations using the standard xml.etree.ElementTree library. This parser is vulnerable to XML External Entity (XXE) attacks, allowing an attacker to potentially read local files or conduct server-side request forgery (SSRF) via malicious XML input.\n- [PROMPT_INJECTION]: The skill exhibits an attack surface for indirect prompt injection through its data ingestion workflows.\n
  • Ingestion points: Reading and parsing XML, JSON, and YAML files in dataset_pipeline_builder.py and vision_model_trainer.py.\n
  • Boundary markers: No delimiters or instructions are used to separate untrusted data from processing logic.\n
  • Capability inventory: The skill has the ability to write files, suggest shell command execution, and contains an unsafe deserialization path through torch.load.\n
  • Sanitization: The scripts use standard parsers without security hardening or validation of the data content.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 2, 2026, 02:36 PM