The Agent Skills Directory

[PROMPT_INJECTION]: No attempts to bypass safety filters or override agent behavior were found in the instructions or metadata.
[CREDENTIALS_UNSAFE]: No hardcoded secrets, API keys, or private credentials were detected. Configuration templates use empty strings or placeholders (e.g., ???) for user input.
[DATA_EXFILTRATION]: No suspicious network activity or attempts to read sensitive system files (like SSH keys or environment variables) were found. Use of external storage paths like S3 are provided as illustrative examples.
[EXTERNAL_DOWNLOADS]: No downloads from untrusted third-party sources. The skill references local tool execution (mal CLI) within a defined NVIDIA TAO container environment.
[COMMAND_EXECUTION]: Commands are restricted to a specific CLI tool within the TAO Toolkit environment. No arbitrary or destructive shell command execution or privilege escalation attempts were found.
[OBFUSCATION]: No hidden or encoded malicious logic was found. The skill.oms.sig file contains standard cryptographic signature metadata (X.509 certificates and DSSE envelopes) used for integrity verification.
[INDIRECT_PROMPT_INJECTION]: While the skill processes external datasets (images and annotations), the interaction is focused on machine learning training and evaluation workflows, with no mechanisms identified that would allow untrusted data to execute commands or influence the agent's logic.

tao-train-mask-auto-label