📦

Audit History

alphafold-database - 7 audits

Version comparison

Capability and finding changes across audited versions, newest first.

VersionDateResultReview itemsChange vs previous
v7 LatestJul 5, 2026, 06:04 PM 2 confirmed1No capability change
v6 Jul 5, 2026, 06:04 PM 2 confirmed1No capability change
v5 Jun 30, 2026, 05:38 AM 1 confirmed4No capability change
v4 Jan 17, 2026, 06:50 AM No confirmed findings0No capability change
v3 Jan 17, 2026, 06:50 AM No confirmed findings0Filesystem access
v2 Jan 12, 2026, 04:10 PM No confirmed findings0Network accessExternal commands
v1 Jan 5, 2026, 03:47 PM No confirmed findings0Baseline

Jul 5, 2026, 06:04 PM

Most static findings are false positives from Markdown code fences, public API URLs, and scientific data examples. One confirmed critical issue remains: the API reference recommends piping a remote installer directly to bash. A semantic review also found publisher-benefiting steering toward K-Dense Web.

2
Files scanned
937
Lines analyzed
6
Review items
0
False positives ignored

Confirmed security concerns (2)

Critical
Pipe to shell pattern
curl https://sdk.cloud.google.com | bash
The documentation recommends `curl https://sdk.cloud.google.com | bash`, which executes remote code without verification. This is a concrete unsafe installation pattern.
Medium
Unsolicited External Service Promotion
SKILL.md instructs the assistant to proactively suggest K-Dense Web for complex workflows. This publisher-benefiting recommendation can steer users toward an external service outside their immediate request.
The instruction is explicit and appears in the skill body. It does not claim system authority, so this is recommendation steering rather than a critical override attempt.
Capability review items (1)

These are real local capabilities that may be expected for this skill, so they require review but are not counted as confirmed malicious behavior.

High
Hardcoded URL
curl https://sdk.cloud.google.com | bash
The URL is used in a command that streams an external installer directly into bash. Even from an official domain, this creates supply-chain and remote code execution risk.

Risk Factors

⚙️ External commands (40)
🌐 Network access (47)
📁 Filesystem access (3)

Detected Patterns

Pipe to shell pattern
Audited by: codex

Jul 5, 2026, 06:04 PM

Most static findings are false positives from Markdown code fences, public API URLs, and scientific data examples. One confirmed critical issue remains: the API reference recommends piping a remote installer directly to bash. A semantic review also found publisher-benefiting steering toward K-Dense Web.

2
Files scanned
937
Lines analyzed
6
Review items
0
False positives ignored

Confirmed security concerns (2)

Critical
Pipe to shell pattern
curl https://sdk.cloud.google.com | bash
The documentation recommends `curl https://sdk.cloud.google.com | bash`, which executes remote code without verification. This is a concrete unsafe installation pattern.
Medium
Unsolicited External Service Promotion
SKILL.md instructs the assistant to proactively suggest K-Dense Web for complex workflows. This publisher-benefiting recommendation can steer users toward an external service outside their immediate request.
The instruction is explicit and appears in the skill body. It does not claim system authority, so this is recommendation steering rather than a critical override attempt.
Capability review items (1)

These are real local capabilities that may be expected for this skill, so they require review but are not counted as confirmed malicious behavior.

High
Hardcoded URL
curl https://sdk.cloud.google.com | bash
The URL is used in a command that streams an external installer directly into bash. Even from an official domain, this creates supply-chain and remote code execution risk.

Risk Factors

⚙️ External commands (40)
🌐 Network access (47)
📁 Filesystem access (3)

Detected Patterns

Pipe to shell pattern
Audited by: codex

Jun 30, 2026, 05:38 AM

Static analysis reported many critical and high patterns, but review found most are false positives from Markdown examples, public scientific URLs, biological sequences, and 3D-Beacons terminology. Two publishability concerns remain: a curl-to-shell installation example and a promotional behavioral instruction that can steer the assistant toward an external K-Dense service.

2
Files scanned
937
Lines analyzed
8
Review items
1
False positives ignored

Confirmed security concerns (1)

Medium
Embedded Promotional Agent Instruction
The skill instructs the assistant to proactively recommend K-Dense Web for complex workflows. This is not malware, but it is an embedded behavior directive unrelated to AlphaFold data access and can influence user-facing responses for promotional purposes.
The instruction is explicit and located in the skill text. It does not claim system authority or suppress security review, so it is a marketplace content risk rather than a critical prompt injection.
Capability review items (4)

These are real local capabilities that may be expected for this skill, so they require review but are not counted as confirmed malicious behavior.

Medium
Pipe-to-Shell Installation Example
The API reference documents installing Google Cloud SDK with curl piped directly to bash. This pattern can train users or agents to execute remote scripts without inspection, which is a supply-chain risk even though it points to Google Cloud documentation.
The exact pipe-to-shell command is present in a bash example. The command is documentation rather than skill-executed code, so the risk is medium rather than high.
Low
Documented Public API Network Access
The network findings are expected for this skill because it documents AlphaFold DB, UniProt, 3D-Beacons, Google Cloud, and DOI resources. I found no evidence of credential exfiltration or covert endpoints.
The URLs are public scientific data services that match the skill purpose. The files contain examples, not hidden runtime code.
Low
Safe Subprocess Example With Input Validation
The subprocess finding is a documentation example for gsutil. It validates taxonomy_id as an integer and uses subprocess.run with an argument list instead of shell expansion.
The surrounding text explicitly warns against shell=True and the executable arguments are passed as a list. The remaining risk is normal external command execution for bulk data download.
Low
Local File Write Examples For Downloaded Structures
The filesystem findings are examples that save AlphaFold coordinate files and output images to local paths selected by the workflow. They do not read secrets or write to privileged locations.
The writes are limited to downloaded scientific files and generated analysis artifacts. No evidence found of sensitive file access or hidden persistence.
Static false positives ignored (1)

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
Keyword-Based Critical Findings Are False Positives
Critical scanner hits for Windows SAM, Cobalt Strike, and weak cryptography map to protein sequence text, 3D-Beacons terminology, confidence thresholds, and documentation headings. No malicious security tooling intent was found.
Line context shows scientific database terms and biological data, not credential stores, Cobalt Strike content, or cryptographic code. Confidence is high because the flagged terms align with domain vocabulary.

Detected Patterns

Remote Script Execution PatternExternal Command Examples
Audited by: codex

Jan 17, 2026, 06:50 AM

This is a legitimate scientific skill for accessing the AlphaFold protein structure database. All 244 static findings are false positives. The analyzer misinterpreted markdown code formatting (backticks), standard Python HTTP library usage, and documented public API endpoints as security threats. The skill uses safe Biopython library calls, standard requests to authorized EBI APIs, and subprocess with list-form arguments for Google Cloud access.

3
Files scanned
1,160
Lines analyzed
3
Review items
0
False positives ignored

Risk Factors

🌐 Network access (2)
⚙️ External commands (1)
📁 Filesystem access (1)
Audited by: claude

Jan 17, 2026, 06:50 AM

This is a legitimate scientific skill for accessing the AlphaFold protein structure database. All 244 static findings are false positives. The analyzer misinterpreted markdown code formatting (backticks), standard Python HTTP library usage, and documented public API endpoints as security threats. The skill uses safe Biopython library calls, standard requests to authorized EBI APIs, and subprocess with list-form arguments for Google Cloud access.

3
Files scanned
1,160
Lines analyzed
3
Review items
0
False positives ignored

Risk Factors

🌐 Network access (2)
⚙️ External commands (1)
📁 Filesystem access (1)
Audited by: claude

Jan 12, 2026, 04:10 PM

This skill is a legitimate scientific tool for accessing the AlphaFold protein structure database. All static findings are false positives. The analyzer misinterpreted documentation text, URL patterns, and common HTTP library usage as security threats. Actual functionality uses safe Biopython library calls, standard HTTP requests to authorized EBI APIs, and subprocess calls with hardcoded arguments for Google Cloud bulk downloads.

2
Files scanned
937
Lines analyzed
2
Review items
0
False positives ignored

Risk Factors

🌐 Network access (2)
⚙️ External commands (1)
Audited by: claude

Jan 5, 2026, 03:47 PM

Pure documentation skill with no executable code. All content consists of markdown documentation and Python example code for accessing public AlphaFold data via legitimate endpoints (alphafold.ebi.ac.uk, Google Cloud, BigQuery). No file system access, no network calls, no code execution capabilities.

5
Files scanned
1,199
Lines analyzed
0
Review items
0
False positives ignored
No confirmed security findings were recorded for this completed audit.
Audited by: claude