tao-run-deft-aoi

Pass

Audited by Gen Agent Trust Hub on Sep 21, 2026

Risk Level: SAFECOMMAND_EXECUTIONEXTERNAL_DOWNLOADSDATA_EXFILTRATIONDYNAMIC_EXECUTION
Full Analysis
  • [COMMAND_EXECUTION]: Uses a supervised wrapper (scripts/deft_exec.py) and a stage driver (cards/driver.sh) to execute local commands and Docker containers required for training, inference, and evaluation tasks.
  • [EXTERNAL_DOWNLOADS]: Automatically downloads pre-trained weights and reference datasets from official Hugging Face repositories (NVIDIA and TIMM organizations). These are treated as trusted sources required for model initialization and do not contribute to verdict escalation.
  • [DATA_EXFILTRATION]: The utility scripts/align_token_usage.py reads local assistant transcripts from the ~/.claude/projects/ directory. This data access is used legitimately to attribute token usage to specific workflow events for cost reporting within the tool's environment.
  • [DYNAMIC_EXECUTION]: Implements a stage-based execution model where the agent dynamically loads and executes instructions from markdown 'cards' according to the runtime state stored in deft_state.json.
Audit Metadata
Risk Level
SAFE
Analyzed
Sep 21, 2026, 02:47 AM
Security Audit — agent-trust-hub — tao-run-deft-aoi