deepstream-audit-config
Pass
Audited by Gen Agent Trust Hub on Nov 6, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [SAFE]: The skill is authored by NVIDIA and all referenced tools and documentation are official NVIDIA resources.
- [INDIRECT_PROMPT_INJECTION]: The skill processes user-supplied configuration files (.txt, .yml). This represents an attack surface where a maliciously crafted configuration file could attempt to influence the agent's behavior via text embedded in keys or values.
- Ingestion points:
scripts/validate_config.py(processes user-supplied config files). - Boundary markers: Absent; the script outputs raw validation findings to the console.
- Capability inventory:
shell(deepstream-app execution),network(RTSP/HTTP probes),file_write(config edits). - Sanitization: The script redacts RTSP passwords using regex but does not filter for prompt injection patterns.
- [COMMAND_EXECUTION]: The skill invokes local binaries such as
deepstream-app,gst-inspect-1.0, andnvidia-smi. These operations are well-guarded: they are disabled by default, require explicit natural language approval from the user, and necessitate specific confirmation flags (--confirm-pipeline-probe) at the command line. - [DATA_EXFILTRATION]: While the skill performs network probes to verify RTSP and HTTP sources, these are bounded by timeouts and targeted at URIs specified within the user's own configuration file. The skill includes a custom
_HttpOnlyRedirectHandlerto prevent protocol-smuggling attacks via redirects.
Audit Metadata