tao-run-inference-service

Warn

Audited by Gen Agent Trust Hub on Sep 17, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONPRIVILEGE_ESCALATIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
  • [PRIVILEGE_ESCALATION]: The skill instructs the agent to prefix container commands with 'umask 0' in the inner command templates.
  • Evidence: Found in SKILL.md Section 4.1 and references/code-templates.yaml for 'cosmos-rl', 'cosmos-predict2.5', and 'tao-dataservices'.
  • Impact: This sets the file mode creation mask to zero, resulting in world-writable and world-readable permissions (0777 equivalent) for all files and directories created by the service, allowing unauthorized local users to access or modify service artifacts.
  • [COMMAND_EXECUTION]: The skill involves extensive construction and execution of shell commands using system tools incorporating user-controlled parameters.
  • Evidence: Python templates in references/code-templates.yaml use the 'subprocess' module to call system binaries including 'ssh', 'squeue', 'sacct', and 'kubectl'.
  • Evidence: SKILL.md Section 4.1 describes building shell command strings for container execution that interpolate user-provided values like 'model_path' and 'host_port'.
  • [DYNAMIC_EXECUTION]: The skill provides numerous Python code templates for the agent to instantiate and execute during runtime.
  • Evidence: Found in references/code-templates.yaml, which contains templates for 'job_payload_builder', 'registry_write', 'readiness_check', and 'stop' flows.
  • Operations: These templates perform sensitive tasks including filesystem writes to '/tmp/tao-inf-ms-state.json', spawning background processes for SSH port-forwarding, and making network requests.
  • [INDIRECT_PROMPT_INJECTION]: The skill processes user-provided inference request bodies and model path values, presenting a vulnerability surface for indirect injection.
  • Evidence: SKILL.md Section 6.2 and references/request.yaml define methods for sending multimodal messages to the microservice.
  • Sanitization: Untrusted data from the user is interpolated into request bodies and command arguments with minimal sanitization beyond basic URI scheme checks for model paths.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Sep 17, 2026, 02:56 AM
Security Audit — agent-trust-hub — tao-run-inference-service