AegisKernel proposes deterministic security controls for autonomous AI agents

Summary

AegisKernel is an open-source repository proposing a zero-trust security runtime for autonomous AI agents that can access files, databases, cloud APIs, and shell terminals. Its design wraps external data in tainted strings, preserves taint through common string operations, and blocks dangerous tool arguments when they lack the required capability grant. The project also describes an interceptor for shell, file-write, and database-modification actions, plus SHA-256 HMAC audit logs intended to make security records tamper-evident. A CrewAI benchmark included in the repository reports four attack classes as exploited in standard CrewAI and blocked with AegisKernel, with listed interception times from 0.016 to 0.061 milliseconds; the project summarizes those results as 0% versus 100% defense. These are repository-provided benchmark claims rather than independently validated results in the supplied material. The package exposes an AegisShield decorator and an integration wrapper for existing CrewAI or LangChain agents, and includes tests, a browser demo, and MIT licensing. The repository labels the project version 0.2.0 and offers paid design-partner services for teams deploying autonomous agents.