neuropixels-analysis
Pass
Audited by Gen Agent Trust Hub on Sep 15, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONMETADATA_POISONING
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data formats such as SpikeGLX, Open Ephys, and NWB files which can contain user-defined metadata. These formats serve as ingestion points for untrusted data that the agent then visualizes and interprets during 'AI-assisted curation'.\n
- Ingestion points: Data loading functions in
scripts/neuropixels_pipeline.pyandSKILL.md(e.g.,si.read_spikeglx,si.read_openephys,si.read_nwb).\n - Boundary markers: The instructions lack specific delimiters or warnings to ignore embedded instructions within data metadata fields.\n
- Capability inventory: The skill allows the agent to execute spike sorters via subprocesses (
si.run_sorter), write files to the local system (rec.save()), and perform network operations to send visual reports to AI model providers (npa.analyze_unit_visually).\n - Sanitization: No sanitization or validation of metadata fields is performed before the content is presented to the agent for visual or text-based analysis.\n- [METADATA_POISONING]: The
skill-authorfield in the frontmatter ofSKILL.mdidentifies the author as 'K-Dense Inc.', which differs from the provided system author context ('leonchaox'). While likely representing a corporate alias, this discrepancy is noted as inaccurate metadata.
Audit Metadata