Wider Security AEGIS

Wider Security AEGIS (AI Execution Guardian and Isolation System) is a security layer designed to supervise and constrain the actions of AI agents at runtime, with the goal of containing the impact of a compromised or manipulated agent. The core problem AEGIS addresses is that traditional guardrails filter model inputs and outputs but do not prevent a hijacked agent from taking harmful real-world actions. Once an agent is compromised via indirect prompt injection or a poisoned tool, it can read data, modify systems, and execute commands if it holds the necessary permissions. AEGIS addresses this through a structural "three-body separation of duties" model: - The Worker: The AI agent itself. It can propose actions but has no ability to execute them directly. - The Supervisor: Evaluates every proposed action through a multi-stage pipeline, including fast deterministic policy checks before any model-based judgment. - The Executor: The only component that can carry out an action, and only those explicitly approved by the Supervisor. Each component runs in a separate, network-isolated container and communicates through a single controlled channel. This architecture assumes agent compromise will occur and structurally limits the blast radius rather than relying on prevention alone. Key operational characteristics: - Fully on-premises deployment with no external cloud calls required, targeting air-gapped or restricted environments - Tamper-evident, hash-chained audit log for every decision, verifiable offline - Controls mapped to the OWASP Agentic Top 10 - Exportable evidence bundles for auditors and accreditors AEGIS is patent pending and currently available to early-access design partners.