Cross-platform installers + license activation + arXiv v2 paper
- macOS Apple Silicon .dmg (cmux-style designed installer), Windows x64 .msi, Linux amd64/arm64 .deb — all on GitHub Releases + downloadable from /download.
- Sign-in with AEGIS device-link flow (Cursor/Warp-style OAuth): Cockpit → browser → aegistraces.com → local gateway callback. No key-paste required for existing paid customers.
- Open-core license enforcement: gateway hot-swaps tier on activation; 24h revalidation with 7-day grace period; deactivation removes tier immediately.
- Stripe subscription flow + Billing Portal: manage subscription without a support ticket.
- arXiv:2603.12621 v2 paper: 3-layer cost-aware cascade (rules + XGBoost + LLM judge) reaches 99.9% block rate at 1.06 ms P50 latency and $0.05 total cost on 5,525-record ToolGuard-Bench. Adversarial ≥ 97.5% under 10 obfuscation transforms.
- CI matrix rebuild: macOS ARM + Windows + Linux all building unsigned in GitHub Actions; workflow_dispatch supports backfill to existing tags.
Full release notes on GitHub →
First desktop release — Cockpit + gateway bundled
- AEGIS Compliance Cockpit shipping as a Tauri desktop app with embedded Node.js gateway + SQLite trace store.
- 14 framework SDKs supported (Anthropic, OpenAI, LangChain, CrewAI, Gemini, Bedrock, Mistral, LlamaIndex, smolagents; Anthropic JS, OpenAI JS, LangChain JS, Vercel AI SDK; Go).
- Tamper-evident audit chain: Ed25519 signatures + SHA-256 hash chain (RFC 6962 style).
- Human-in-the-loop pending queue: SDK suspends agent execution on pending decision; reviewer approves in the Cockpit within one polling cycle.
- 22 detection patterns across 7 categories (SQL injection, path traversal, shell injection, prompt injection, sensitive files, data exfiltration, PII leakage).
Full release notes on GitHub →
arXiv v1 demo paper — first public release
- arXiv:2603.12621 v1: "No Tool Call Left Unchecked — a Pre-Execution Firewall and Audit Layer for AI Agents".
- Initial rules-only defense evaluated on a 48-attack curated suite (100% blocked, 1.2% false positive, 8.3 ms median latency).
- Open-source engine + Cockpit UI + demo scenarios (SQL injection interception, path traversal, human-in-the-loop approval).
- Note: v1 rules-only design blocked 48/48 curated but only 1.2% on the larger 5,525-record ToolGuard-Bench introduced in v2 — motivating the 3-layer cascade.
Full release notes on GitHub →