Versioned security assessment

Report ID: SA-F95C1B0C

6/30/2026, 5:52:45 AM

biomni security assessment v5

Skill Security Certification Report

Audit History
Audit model: codex Historical report
Skill name
biomni
Version
v5
Maintainer
K-Dense-AI
Source
github.com/K-Dense-AI/claude-scientific-skills/tree/main/scientific-skills/biomni
Coverage
6 Files scanned · 2,865 Lines analyzed
Policy version
Unavailable

Confirmed finding summary

No confirmed security findings

The completed audit recorded no confirmed security findings. This is not proof that the Skill has no side effects.

Installation context

Historical evidence

This report may not describe the currently installable artifact. Open the current Skill page for install guidance.

Open current Skill page

This report does not block or authorize the manifest or ZIP.

Static analysis produced many hits, and most Markdown backtick, weak-crypto, C2, and hardcoded-URL detections are false positives from documentation examples. However, the skill intentionally supports autonomous LLM-generated code execution, external MCP tools, local file writes, and API key storage, so it should not be published without strong warnings and sandbox guidance.

Report position

Historical report

Open audit history before using this report to install.

Audit attestation

Not attestable

The required immutable binding is incomplete.

Human verification

Not verified

No human verification is recorded for this report.

Coverage

6 Files scanned · 2,865 Lines analyzed

6 items shown for review

Limitations

This report does not claim runtime or sandbox execution and does not prove the absence of side effects.

Evidence chain

Follow the evidence from source binding to the install contract. Available evidence supports verification; it is not a safety guarantee.

  1. Source

    Binding unavailable

  2. Artifact

    Identity incomplete

  3. Audit

    Complete

  4. Install contract

    Open manifest to verify

    Open manifest

Capabilities observed

Observed means this report recorded supporting evidence. Not recorded does not prove that a capability is absent.

Contains scripts

May execute code included with the Skill.

Observed in 1 evidence location

Network access

May connect to external services.

Observed in 18 evidence locations

Filesystem access

May read or write local files.

Observed in 6 evidence locations

Env variables

May read values from the process environment.

Observed in 42 evidence locations

External commands

May invoke commands or programs outside the Skill.

Observed in 251 evidence locations

Capability review items (6)
High
SKILL.md:19-21SKILL.md:101-119SKILL.md:231-236references/api_reference.md:142-159
Autonomous Unsandboxed Code Execution
The skill describes Biomni as generating and executing analysis code, with code execution enabled by default and sandbox mode disabled unless configured. This is a real high-risk capability because LLM-generated biomedical workflows can execute with the user process privileges.
The documentation explicitly states that Biomni generates and executes code and warns that it runs with full system privileges. This confirms the static execution findings as a real security concern, although not malicious intent.
High
scripts/setup_environment.py:82-122scripts/setup_environment.py:125-160scripts/setup_environment.py:293-306
Interactive Credential Collection and Persistence
The setup helper prompts users for LLM provider API keys and can write them into a local .env file. This is legitimate setup behavior, but it increases exposure if run in shared directories or logs are mishandled.
The code directly prompts for named API keys and writes collected values to .env or prints shell export commands. The behavior is clear and intentional, but it is configuration support rather than exfiltration.
High
SKILL.md:189-201references/api_reference.md:333-359references/api_reference.md:362-378
External MCP Command and Network Tool Integration
The skill documents MCP server configuration that can run external commands and pass API keys into network-enabled tools. This expands the execution and network surface when users enable MCP integrations.
The evidence shows documented command-based MCP servers and API key environment mappings. This is optional and user-configured, so confidence is high for risk surface but lower for immediate exploitability.
Medium
scripts/setup_environment.py:22-32scripts/setup_environment.py:47-69scripts/setup_environment.py:199-220
Subprocess-Based Setup and Test Execution
The setup script runs conda commands and can create then execute a generated Biomni test script. The subprocess arguments are mostly fixed, so command injection evidence was not found, but running the test can trigger package behavior, data access, and LLM calls.
The subprocess.run calls are real and the generated test file is executed. The arguments are not shell-concatenated user input, which reduces injection likelihood.
Medium
scripts/generate_report.py:111-185scripts/generate_report.py:197-254scripts/generate_report.py:343-361
Unescaped HTML Report Rendering
The report generator inserts conversation content and titles into HTML without escaping. If untrusted conversation history is converted to HTML and opened in a browser, embedded markup may execute or alter the report.
The converter wraps raw message content into HTML tags without escaping. Exploitability depends on the report input source and how users open generated HTML.
Medium
scripts/setup_environment.py:125-149scripts/setup_environment.py:180-184scripts/setup_environment.py:211-220scripts/generate_report.py:280-299
Local Filesystem Writes for Reports, Examples, and Secrets
The skill writes reports, example scripts, data directories, test files, and .env credential files. These writes are user-facing and mostly expected, but they can expose sensitive research data or credentials if paths are shared.
The file writes are explicit and legitimate for setup and reporting. The main risk is accidental disclosure of credentials or biomedical data, not hidden persistence.

Risk findings

Confirmed security concerns are separated from items that still need review.

No confirmed security findings were recorded for this completed audit.

Expert evidence

Immutable subject identity, scanner metadata, dismissed matches, and source-level evidence.

Artifact subject

Marketplace commit
Unavailable
Content hash
Unavailable
Tree hash
Unavailable
Skill path
Unavailable
Audit payload hash
Unavailable

Analysis metadata

Audit model: codex

Analysis state: Complete

Scope is limited to the recorded files, lines, methods, and evidence. No runtime or sandbox execution is claimed.

Static false positives ignored (3)
Low
SKILL.md:40-65references/api_reference.md:11-20references/use_cases.md:24-48references/llm_providers.md:25-31
Markdown Code Examples Misclassified as Shell Execution
Most external command detections in SKILL.md and references are Markdown fenced code examples, inline model names, or biomedical examples. I did not find evidence that these documentation snippets execute by themselves.
The cited lines are fenced documentation examples rather than Ruby backtick execution. Similar static hits across reference files follow the same Markdown pattern.
Low
SKILL.md:1-3references/use_cases.md:189-199references/use_cases.md:568scripts/generate_report.py:1-5
Biomedical Text Misclassified as Crypto or C2 Indicators
Static blocker hits for weak cryptography and C2 keywords appear to match biomedical abbreviations, model identifiers, SMILES strings, or example task text. No evidence found of cryptographic implementation, command-and-control behavior, or malicious networking.
The inspected context shows domain content and documentation, not cryptographic code or C2 logic. This supports dismissing those static blocker patterns as false positives.
Low
references/llm_providers.md:112-126references/llm_providers.md:253-285SKILL.md:271-276scripts/setup_environment.py:342-344
Hardcoded URLs Are Documentation and Provider Links
Hardcoded URLs point to provider consoles, project documentation, GitHub resources, and local example endpoints. No evidence found that the skill silently sends data to these URLs.
The URLs are visibly documentation, provider setup, or localhost examples. Network access can occur when users configure providers, but these lines do not show covert exfiltration.

Verify and export

JSON report CITATION.cff SARIF 2.1.0 Signed installation manifest

The manifest and lockfile bind install artifacts to cryptographic hashes. This integrity claim is separate from the security assessment.

Audit attestation: not_attestable