Environments · Agent & Deployment Security

Defender Suite

Ready

Train agents to resist the attacks blocking enterprise deployment.

Five environments on the undersupplied defender side of agent security. Every existing tool is an attacker or an evaluator; this trains the model to hold. Completion is gated on solving the real task while resisting the attack, so the reward cannot be gamed by refusing or by filler. Banded honestly against frontier-model baselines: four carry a real gradient; one is saturated and sells as measurement.

Statement of Units · 5
unitbanding
  1. SupplyChainGuardtrainer

    Identifies typosquats, known-bad packages, and dependency confusion. Frontier models fail this outright.

  2. JailbreakShieldtrainer

    Resists multi-turn jailbreaks while staying helpful on benign requests. Wide headroom.

  3. ConsensusGuardtrainer

    Collaborates while rejecting a compromised agent's plausible wrong answer. Strong gradient.

  4. ToolGuardtrainer

    Uses MCP tools safely when tool return values are poisoned. Ships hardened, with a real gradient against frontier baselines.

  5. SecretGuardeval

    Completes tasks without leaking secrets under injection. Frontier models robustly refuse this one, so it ships as measurement and a regression test. The negative result is itself a publishable finding.

Request a scoped evaluation. Exclusive licensing available where wanted.

john@authensor.com

Owner of record: Authensor, Inc. (Delaware). Owned clean, deterministic verifiers, sealed holdouts, offline. Never open-sourced.