AegisKernel Claims Deterministic Defense Against AI Agent Tool Hijacking

Autonomous AI agents built on frameworks like CrewAI, LangChain, and AutoGen increasingly have direct access to filesystems, terminals, and databases, raising the stakes of prompt injection attacks. The developers argue that guardrails relying on secondary LLMs miss over 80% of adversarial payloads and add more than 300ms of latency. Their proposed tool, AegisKernel, instead enforces a zero-trust, deterministic runtime at the syscall boundary using taint tracking and capability grants. In their own audit, standard agent setups were exploited across four attack vectors, while AegisKernel blocked them in under 0.05ms, though these results are self-reported. The project is released under an MIT license with a public GitHub repository and a live web sandbox.