agentdb

AI Security Review

Evidence context: threat category unknown malware; evidence quality weak.

AgentDB v3 presents a critical risk score primarily due to a massive volume of malware signature findings (476) and network indicators (785). The scanner detected 2 secret findings, 2 network activity findings, and 5 code-smell hits, but found zero instances of tool poisoning. In the MCP ecosystem, this pattern—high malware counts with zero tool poisoning—strongly suggests bundle contamination. The 476 malware signatures likely originate from minified JavaScript dependencies or utility libraries (e.g., specialized math/DB kernels) being misidentified by YARA rules. Similarly, the 785 IoCs are almost certainly the 'IPv6 fragment' or 'property chain' false positives typical of obfuscated code. However, the 'unknown' status of the findings forces a cautious stance: without verification, we cannot distinguish between aggressive false positives and a sophisticated supply-chain delivery (e.g., 'Gupti' malware) inside the dist bundle. The recommendation is manual review of the dist/ build artifacts or runtime sandboxing.

Key Reasons

• 476 malware signatures in dist/ likely false positives from minified deps
• Zero tool-poisoning findings despite high-risk classification
• High IoC count (785) consistent with property-access false positives
• Unknown findings preventing automated clearance

False Positive Considerations

• Minified/Bundled Dependencies
• Generic YARA Malware Signatures
• IPv6/Property-Access IoC Extraction Noise

Reviewed 2026-04-12; recommended action: reanalyze; model confidence 60%.