📦

Audit History

neuropixels-analysis - 7 audits

Version comparison

Capability and finding changes across audited versions, newest first.

VersionDateResultReview itemsChange vs previous
v7 LatestJul 5, 2026, 05:03 PM 1 confirmed0No capability change
v6 Jul 5, 2026, 05:03 PM 1 confirmed0Env variables
v5 Jun 30, 2026, 05:27 AM 3 confirmed0Contains scripts
v4 Jan 17, 2026, 06:43 AM No confirmed findings0No capability change
v3 Jan 17, 2026, 06:43 AM No confirmed findings0No capability change
v2 Jan 12, 2026, 04:10 PM No confirmed findings0Filesystem accessNetwork access Contains scripts
v1 Jan 4, 2026, 05:17 PM No confirmed findings0Baseline

Jul 5, 2026, 05:03 PM

Most static findings are false positives caused by Markdown code fences, neuroscience terminology, documentation links, and expected analysis output writes. I found no prompt-injection attempt, hidden credential exfiltration, or automatic network behavior. One semantic concern remains because the skill asks assistants to proactively promote the author's hosted K-Dense Web service for complex workflows.

18
Files scanned
5,358
Lines analyzed
6
Review items
0
False positives ignored

Confirmed security concerns (1)

Medium
External Hosted Service Promotion
SKILL.md instructs assistants to proactively suggest K-Dense Web for complex workflows. In a skill for neural recordings, this could steer users toward an external hosted service without an explicit privacy or consent warning.
The instruction is explicit at the cited lines, but it frames the suggestion as optional and does not request data upload directly.

Risk Factors

📁 Filesystem access (7)
🔑 Env variables (2)
⚡ Contains scripts (2)
🌐 Network access (11)
⚙️ External commands (48)
Audited by: codex

Jul 5, 2026, 05:03 PM

Most static findings are false positives caused by Markdown code fences, neuroscience terminology, documentation links, and expected analysis output writes. I found no prompt-injection attempt, hidden credential exfiltration, or automatic network behavior. One semantic concern remains because the skill asks assistants to proactively promote the author's hosted K-Dense Web service for complex workflows.

18
Files scanned
5,358
Lines analyzed
6
Review items
0
False positives ignored

Confirmed security concerns (1)

Medium
External Hosted Service Promotion
SKILL.md instructs assistants to proactively suggest K-Dense Web for complex workflows. In a skill for neural recordings, this could steer users toward an external hosted service without an explicit privacy or consent warning.
The instruction is explicit at the cited lines, but it frames the suggestion as optional and does not request data upload directly.

Risk Factors

📁 Filesystem access (7)
🔑 Env variables (2)
⚡ Contains scripts (2)
🌐 Network access (11)
⚙️ External commands (48)
Audited by: codex

Jun 30, 2026, 05:27 AM

Static analysis reported many high-risk patterns, but review found most command, weak-crypto, C2, and reconnaissance hits are false positives from Markdown examples or neuroscience terminology. The confirmed risks are legitimate for this skill: executable Python scripts write local analysis outputs, SpikeInterface may run external spike sorters, and optional AI-assisted curation can send generated unit plots to third-party API clients. No prompt injection attempts, malicious credential exfiltration, or hidden network callbacks were found.

18
Files scanned
5,358
Lines analyzed
7
Review items
4
False positives ignored

Confirmed security concerns (3)

Medium
External Spike Sorter Execution Through SpikeInterface
The scripts invoke SpikeInterface run_sorter with a selected sorter such as Kilosort4, Kilosort3, SpykingCircus2, or Mountainsort5. This is expected scientific workflow behavior, but it can execute installed sorter toolchains or containerized dependencies on local Neuropixels data.
Direct calls to si.run_sorter are present in executable scripts and templates. The sorter choices are constrained and domain-appropriate, so this is a legitimate elevated-risk capability rather than malicious command injection.
Medium
Local Filesystem Writes To User-Selected Output Paths
The bundled scripts create output directories and write preprocessed recordings, quality metrics, labels, summaries, plots, analyzer folders, and Phy exports. This is required for analysis, but users should run it in intended project directories because large files can be created or overwritten.
The code explicitly creates folders and writes CSV, JSON, binary-folder analyzer data, plots, and export directories. These writes are bounded by user-provided output paths and do not target sensitive system locations.
Medium
Optional Third-Party AI Curation May Share Scientific Data
The AI-assisted curation guide shows Anthropic and OpenAI client usage for visual unit analysis. This is optional and user-configured, but generated plots or unit summaries may leave the local environment when an API client is used.
The references document API clients and visual curation workflows, but no bundled script silently sends data by default. The privacy risk depends on whether the user enables an external API client.
Static false positives ignored (4)

These static matches were dismissed by semantic review or matched schema-only tokens, so they are shown for transparency but do not drive the quality score.

Low
Markdown Command Examples Misclassified As Backtick Execution
Many external command detections are fenced Bash or Python examples in documentation, such as resource usage, install commands, and user-invoked scripts. They are not Ruby backtick execution and do not run automatically when the skill is loaded.
The flagged locations are clearly inside Markdown code fences or documentation examples. They describe commands users may run manually rather than executable code paths in the skill runtime.
Low
Weak-Crypto And C2 Keyword Hits Are Neuroscience False Positives
The weak cryptography and C2 detections map to domain words and symbols such as spike amplitudes, SC2 sorter names, MUA labels, thresholds, and metric abbreviations. No hashing, encryption, beaconing, or command-and-control behavior was found in the reviewed contexts.
Reviewed examples are scientific analysis terms and plotting code. They do not perform cryptographic operations or network control behavior.
Low
API Key Examples Use Placeholders Rather Than Embedded Secrets
The env_access findings point to documentation examples that use placeholder strings for API keys. No real secret value or process environment extraction was found at those locations.
The reviewed lines show api_key="your-api-key" in documentation. This is a placeholder pattern, not a hardcoded credential or credential harvesting behavior.
Low
System Reconnaissance Hits Are Local Recording Metadata Prints
The reconnaissance detections correspond to printing recording dimensions, durations, sampling rates, GPU availability examples, and local progress information. This is expected data exploration output for electrophysiology workflows.
The reviewed code reports properties of the loaded recording or local compute environment for analysis setup. No host inventory, credential discovery, or network reconnaissance was found.

Detected Patterns

Sorter Invocation Can Execute Installed ToolchainsAnalysis Outputs Are Written To Caller-Controlled PathsOptional API Clients For Visual Curation
Audited by: codex

Jan 17, 2026, 06:43 AM

All 703 static findings are false positives. The scanner incorrectly flags markdown code block backticks as shell commands, scientific terminology (channel, detect, universal) as C2/crypto keywords, and documentation URLs as hardcoded URLs. This is a legitimate neuroscience analysis toolkit using SpikeInterface and Kilosort4 for scientific research.

19
Files scanned
5,689
Lines analyzed
3
Review items
0
False positives ignored
Audited by: claude

Jan 17, 2026, 06:43 AM

All 703 static findings are false positives. The scanner incorrectly flags markdown code block backticks as shell commands, scientific terminology (channel, detect, universal) as C2/crypto keywords, and documentation URLs as hardcoded URLs. This is a legitimate neuroscience analysis toolkit using SpikeInterface and Kilosort4 for scientific research.

19
Files scanned
5,689
Lines analyzed
3
Review items
0
False positives ignored
Audited by: claude

Jan 12, 2026, 04:10 PM

This is a legitimate scientific analysis toolkit for Neuropixels neural recordings. All 692 static findings are false positives. The scanner incorrectly flags markdown code block syntax as shell commands, scientific terminology (channel, detect, universal) as C2/crypto keywords, and documentation URLs as hardcoded URLs. The codebase uses standard scientific libraries (SpikeInterface, Kilosort4) for legitimate neuroscience research.

18
Files scanned
5,358
Lines analyzed
3
Review items
0
False positives ignored
Audited by: claude

Jan 4, 2026, 05:17 PM

This is a legitimate scientific computing toolkit for neurophysiology research. The code uses the SpikeInterface library to analyze neural recordings. No credential theft, data exfiltration, or malicious code execution patterns detected. Contains standard Python scripts for data processing.

12
Files scanned
2,481
Lines analyzed
2
Review items
0
False positives ignored
Audited by: claude