aegis
Sign in Start free

Public roadmap · updated 2026-09

AEGIS roadmap

What's shipped, what's landing this quarter, what's planned next. Public because operators of regulated deployments need to plan around it. Cross-references /changelog (shipped detail), GitHub Releases, and arXiv:2603.12621 for the research anchor.

Q3 2026

✓ Shipped

  • arXiv v2: 3-layer cost-aware cascade paper

    Cascade design, an audit of the benchmark we had used to validate it, and a corpus rebuilt by running a live agent. Content patterns catch 0.4% of real agent-emitted attacks; instruction provenance reaches AUC 0.991 against a control where benign calls also draw on tool output.

  • Cross-platform installers (macOS ARM + Windows + Linux)

    GitHub Actions matrix build → GitHub Releases. Cmux-style designed .dmg for macOS; .msi for Windows; .deb for Linux amd64 + arm64.

  • Sign-in with AEGIS device-link flow

    Cursor/Warp-style OAuth: Cockpit → default browser → aegistraces.com → local gateway callback. No manual license-key paste.

  • Open-core license enforcement

    Gateway hot-swaps tier on activation; 24h revalidation with 7-day grace period. Stripe subscription + Billing Portal self-serve.

Q4 2026

⟳ In progress

  • macOS + Windows code-signing

    Apple Developer Program ($99/yr) + Windows EV certificate ($260/yr) → signed installers, no Gatekeeper / SmartScreen warnings on first launch.

  • ISO 27001 evidence pack

    Enterprise-tier addition to existing SOC 2 Type II. Ships as a signed JSON pack alongside the EU AI Act Article 12–15 evidence.

  • Multi-lingual PII + content-safety detectors

    Expand the current English-centric detector catalogue to cover the top 10 EU + APAC languages. Feeds into the L1 rules and L2 XGBoost features.

  • MCP audit server (v1)

    AEGIS gateway ships an MCP endpoint at /mcp-audit that other Claude / GPT agents can call to inspect their own trace history. Dogfooding surface.

  • Agent behavioural baseline v2

    Per-agent z-score anomaly detection over the L2 feature space (already shipping); v2 adds delegation-aware baselining across multi-agent handoffs.

Q1 2027

◇ Planned

  • Learning-based L2 fine-tuning UI

    Let operators re-train the L2 XGBoost classifier on their own labelled traces without leaving Cockpit. Model persistence + policy versioning.

  • Reasoning-chain verification (v1)

    Consistency check between the LLM's chain-of-thought and its actual tool call — a new detector class that catches "the model said A, then called B" divergence.

  • Multi-agent cascade analysis

    Track risk propagation when one agent's output becomes another's input. First-class delegation IDs already ship; cascade-level risk scoring is the next layer.

  • Adaptive trust scoring

    Adjust per-agent approval thresholds automatically based on behavioural history. Requires the baseline v2 to land first.

Follow along

Full changelog Atom feed JSON Feed GitHub Releases