Agencia Guedes

AI Security Review

Evidence context: threat category none; evidence quality moderate.

Agencia Guedes is a Chrome extension that layers CRM features onto WhatsApp Web, published from [email protected]. Version 7.4.3.87 produced 25 findings, every one of them medium severity: 22 network detections, 2 obfuscation detections and a single manifest-analysis entry. Nothing matched a malware signature, a threat-intel indicator, a hardcoded credential, a dependency alert or a code-smell rule.

The network findings name the libraries the scanner recognised, not destinations. NET-SOCKET_IO fires in content/assets/js/v_7_4_3_87_0a650f0a-57a6-4989-860e-3a9d8748d28966.js and ...12.js, NET-JQUERY_AJAX in ...23.js and ...24.js, and NET-FETCH in roughly a dozen further chunks plus background.js:137. Socket.IO is a real-time transport, and a CRM that mirrors WhatsApp conversations across devices needs one. The scan extracted no endpoint list, so there is no domain to weigh here, only the presence of those library calls inside bundled files.

Both obfuscation hits read as build output. OBFUSCATION-LARGE_BASE64 sits in funnil/assets/js/9cf42769-2ad5-4714-8614-9a2b4c455349.CgnacD38.js, a filename carrying a content hash and a short build suffix, the shape a bundler gives a chunk, and a long base64 block in that position normally carries an inlined image, font or source map. OBFUSCATION-FUNCTION_INDIRECT in content/assets/js/v_7_4_3_87_0a650f0a-57a6-4989-860e-3a9d8748d289260.js flags indirect calls, which is what a minifier leaves behind when it rewrites a direct call into a bracketed property lookup. Neither pattern hides a payload; both are ordinary in shipped JavaScript.

The chunk names support that reading. Files such as v_7_4_3_87_0a650f0a-57a6-4989-860e-3a9d8748d289137.js embed the version string 7.4.3.87 and one shared build UUID, numbered sequentially. That is a flat chunked bundle from a repeatable build, and it explains why a nine-user extension generates two dozen network findings: the scanner reports once per chunk instead of once per library.

The strongest case against this reading is that a WhatsApp CRM sits on top of private conversations and ships a live socket transport, so chat content could in principle be pushed to a server that never appeared in the scan. True, though the scan only shows what the code can do, not what it does. The three signals that would confirm that concern, a named non-CDN endpoint, a threat-intel match, or code written to package and ship message data, are all absent. The manifest detail captured in this pass is also thin: no permission strings were recorded, which is a gap in the scan rather than in the extension. A re-run that captures the manifest permissions and resolved endpoints would show whether the socket connection targets the vendor's own service.

On this evidence the findings describe what the extension is built from.

Key Reasons

• All 22 network findings are library detections (NET-SOCKET_IO, NET-JQUERY_AJAX, NET-FETCH) inside bundled chunk files, with no endpoint or domain extracted.
• Zero IoC, malware-signature, secret, dependency and code-smell findings across the whole scan.
• OBFUSCATION-LARGE_BASE64 sits in a content-hashed bundle chunk (funnil/assets/js/9cf42769-2ad5-4714-8614-9a2b4c455349.CgnacD38.js), the expected position for an inlined asset.
• OBFUSCATION-FUNCTION_INDIRECT in content/assets/js/v_7_4_3_87_0a650f0a-...260.js is a standard minifier artefact with no accompanying payload.
• Chunk filenames embed version 7.4.3.87 and one shared build UUID, consistent with an automated build pipeline rather than hand-written loader code.

False Positive Considerations

• Network detections name libraries, not destinations
• Base64 and indirect-call obfuscation hits in bundler output
• Per-chunk reporting multiplies one library signal into dozens of findings
• No IoC, malware-signature or code-smell corroboration anywhere in the scan

Reviewed 2026-10-12; recommended action: monitor; model confidence 72%.