The Agent Skills Directory

[PROMPT_INJECTION]: The skill contains technical documentation and answering rules that establish clear boundaries for the agent's behavior. No attempts to override safety guidelines, bypass content filters, or extract system prompts were found.
[DATA_EXFILTRATION]: No network operations (such as curl or wget) or access to sensitive local file paths (such as .ssh or .aws) were detected. Mentions of external platforms like HuggingFace and GitHub are limited to official vendor repositories and model hosting locations.
[REMOTE_CODE_EXECUTION]: The skill does not contain instructions for downloading and executing remote scripts. While it mentions various training and evaluation SDKs (vLLM, NeMo Evaluator), these are described as technical specifications of the research report rather than runtime commands for the agent.
[OBFUSCATION]: No evidence of encoded content, hidden characters, homoglyphs, or other obfuscation techniques designed to conceal malicious intent was found in any of the analyzed files.
[CREDENTIALS_UNSAFE]: No hardcoded API keys, tokens, or private secrets were found. References to environment variables are absent, and the skill focuses on public research data.
[DYNAMIC_EXECUTION]: The skill is composed entirely of static Markdown and metadata. There is no use of dynamic code generation, unsafe deserialization, or runtime script execution.
[COMMAND_EXECUTION]: No shell commands or subprocess spawning patterns were identified. CLI tools mentioned in the documentation (e.g., Slurm, Ray) are part of the technical description of the model's training infrastructure and are not invoked by the skill itself.

nemotron-ultra