Version 0.4.1 — September 2026

What's New

v11b Neural Threat Detection Model — Lens's on-device Char CNN-BiLSTM model has been upgraded from v9 to v11b, matching the AegisGate Platform and Rampart models. This brings improved adversarial prompt injection detection with the same ~5-50ms inference speed — 100% in your browser, zero data sent anywhere.

Model weights re-exported as float16 gzip base64 JSON (3.75 MB)
Latin-1 encoding fix: inline encoder now correctly handles characters 128-255
keyWalkReverse evasion-resistance mapping ported from Platform and Rampart
Adversarial Evasion Suite — Added a 550-test adversarial evasion suite (50 evasion transforms x 11 MITRE ATLAS attack payloads) to validate detection resistance against techniques like character substitution, case manipulation, punctuation injection, and encoding tricks.

Bug Fixes

Fixed stale comment in char-normalizer.js
Fixed MODEL_VERSION assertion in performance tests
Fixed CI: smoke tests now trigger on push to main (previously only ran on PRs)
Security and Privacy

100% on-device — no prompt text, URLs, or page content ever leaves your browser
Zero external dependencies — no npm packages, no CDNs, no telemetry
SHA-256 model weight verification — ML model is cryptographically verified before loading to prevent supply-chain tampering
176 regex patterns plus ML — detects SSN, credit cards, API keys, passwords, AWS credentials, GitHub tokens, and adversarial prompt injection attacks
10 AI providers — ChatGPT, Claude, Gemini, Copilot, DuckDuckGo, Perplexity, Mistral, Grok, DeepSeek, Meta AI
Fail-visible — every detection is shown to you; you decide to cancel, edit, or proceed
About This Release

AegisGate Lens is a free, open-source browser extension (Apache 2.0) that catches sensitive data and prompt injection attacks before they reach AI services. Think of it as a spellchecker for privacy — but instead of catching typos, it catches your Social Security number, API keys, and passwords before you hit send.

No account required. No upsell. No data collection. Free. Forever.

🔒 Security Audit Remediation Release

v0.3.2 resolves all 27 findings from a comprehensive third-party security audit (1 CRITICAL, 5 HIGH, 10 MEDIUM, 11 LOW). This is a security-hardening release with no new features.

• Model Integrity Verification: ML threat detector now verifies SHA-256 hash of model weights before loading, preventing supply-chain tampering

• ReDoS Prevention: Fixed catastrophic backtracking in international phone number and XSS mutation regex patterns
• Path Traversal: Build script now validates all file paths stay within output directory
• SSRF Prevention: Backend URL validation requires HTTPS and warns on private/loopback IPs

• XSS Prevention: All dynamic content in detection banners is now HTML-escaped
• Kill Switch: Fixed kill switch to disable extension immediately (was checked too late)
• Event Listener Cleanup: Properly removes event listeners to prevent memory leaks
• Input Length Cap: Added 50,000 character limit to prevent performance degradation
• CI Security: Pinned all 16 GitHub Actions to specific versions, fixed shell injection vulnerability

• Privacy: Removed raw prompt text from global window object (now stores length only)
• Zero-Width Character Evasion: Strips invisible Unicode characters before detection
• Pattern Cleanup: Removed overly-broad medical billing code patterns that matched any 5-digit number

• ✅ 518/518 tests pass
• ✅ OPSEC scan: 0 failures
• ✅ semgrep OWASP+audit: 0 findings
• ✅ No external npm dependencies (vanilla JavaScript only)

All detection happens on-device. No data leaves your browser.

For full details: https://github.com/aegisgatesecurity/aegisgate-lens/blob/main/CHANGELOG.md

• 23 New SOC Detection Patterns — SWIFT/BIC banking codes (3), CPT/HCPCS medical billing codes (11), and OT/ICS protocol patterns (9: Modbus, DNP3, OPC-UA). Brings total to 155 regex patterns.
• Firefox MV3 Support — Firefox 142+ via Manifest V3 with browser_specific_settings.gecko.
• Browser compatibility shim — src/browser-compat.js aliases browser.* to chrome.* for cross-browser support. No-op on Chrome.
• Dynamic injection fix — Background.js content script file list corrected to match manifest.
• 530 tests — 518 passing + 12 skipped (ML perf tests). All green.

AegisGate Lens is a privacy-first browser extension that detects sensitive data in AI chat prompts before you hit send. It runs 100% on-device — no prompt text, URLs, or page content ever leaves your browser.

When you type into an AI chat (ChatGPT, Claude, Gemini, Copilot, etc.), Lens scans your prompt in real-time (debounced 250ms) and shows a banner if it detects:

1. PII — SSN, email, phone, credit cards, medical record numbers, CPT/HCPCS codes, driver's license, passport, bank accounts, SWIFT/BIC codes, international IDs (69 patterns)
2. Secrets — API keys (AWS, GitHub, OpenAI, Stripe, Slack), OAuth tokens, private keys, database credentials (41 patterns)
3. XSS — Cross-site scripting payloads (12 patterns)
4. Compliance — OWASP LLM Top 10, MITRE ATLAS, EU AI Act, NIST CSF, ISO 27001, CCPA, LGPD, PIPEDA, POPIA (24 patterns)
5. OT/ICS Protocols — Modbus, DNP3, OPC-UA control manipulation (9 patterns)
6. ML Threat — Adversarial prompt injection detection via Char CNN-BiLSTM neural network running in pure JavaScript (1 model, ~5-50ms)

The banner lets you Cancel send, Edit manually, or Send anyway. You can also dismiss false positives for 24 hours.

• Zero telemetry by default. No prompt text, no URLs, no page content, no PII, no keystroke timing, no mouse movement, no session IDs, no IP addresses ever collected.
• 100% on-device processing. All detection runs in the browser content script.
• Opt-in anonymous telemetry only. If the user explicitly enables it, only hashed domain names and rounded timestamps are sent — never prompt content.
• 12 privacy non-negotiables documented in docs/PRIVACY-POLICY.md and docs/SECURITY.md.
• data_collection_permissions in the manifest declares "required": ["none"] — no data collection is required for the extension to function.