dicom-series-preflight
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes metadata from user-supplied DICOM files. An attacker could theoretically place malicious instructions within DICOM header fields (e.g., 'SeriesDescription' or 'PatientComments') to attempt to influence an agent that subsequently reads the generated JSON report.
- Ingestion points:
scripts/preflight_series.pyrecursively scans and reads headers from all DICOM files in the provided directory. - Boundary markers: Extracted data is structured into a JSON report based on
validators/output_schema.json. - Capability inventory: The skill is permitted to use the
Bashtool and performs file system reads and writes as described in the manifest. - Sanitization: Metadata values are converted to strings using the
_safe_strfunction, though no specific filtering for prompt injection markers is performed.
Audit Metadata