physicalai-train-training-a-policy
Pass
Audited by Gen Agent Trust Hub on Oct 12, 2026
Risk Level: SAFE
Full Analysis
- [PROMPT_INJECTION]: No patterns of instruction overriding, safety bypass, or role-play injections were detected in the instructions or metadata.
- [DATA_EXFILTRATION]: The skill references local project directories (e.g., experiments/, library/configs/) and uses official Hugging Face repository identifiers (e.g., lerobot/pusht), which is standard behavior for machine learning workflows. No unauthorized network operations or credential exposure were found.
- [REMOTE_CODE_EXECUTION]: The command execution is limited to the physicalai CLI tool and standard testing utilities like pytest. No untrusted remote scripts are downloaded or executed.
- [OBFUSCATION]: No obfuscated strings, encoded commands, or hidden characters were identified.
- [PRIVILEGE_ESCALATION]: The skill does not request or attempt to acquire elevated system privileges; all operations are conducted within the user's environment.
- [INDIRECT_PROMPT_INJECTION]: While the skill interacts with dataset repositories, there is no evidence of processing untrusted data in a way that would influence agent behavior or execute embedded commands.
- [DYNAMIC_CONTEXT_INJECTION]: No load-time dynamic execution patterns (e.g., shell command interpolation) were found in the markdown.
Audit Metadata