Author(s): Diogo Santos

Originally published on Towards AI.

Your agent has 1,000 tools and no idea which ones it’s allowed to call. agent-kernel gives every tool call an HMAC capability token, a policy gate, and a tamper-evident audit trail — in-process, in two dependencies.

Your agent works. You wired up an LLM, handed it a pile of MCP tools, and watched it plan, call list_invoices, summarize the result. Demo-day magic.

The article argues that most agent frameworks are unsafe in practice because tool registries are effectively flat, tool outputs flow straight back into the prompt, and logs are often insufficient to answer “who did what, and under whose authority.” It presents agent-kernel (weaver-kernel) as a small Python security kernel that sits between an LLM loop and tools, enforcing capability-based authorization for every tool call and producing a tamper-evident, hash-chained audit trail. It explains the end-to-end lifecycle (register → grant → invoke → expand → explain), detailing how HMAC-signed, time-bounded capability tokens prevent confused-deputy attacks via binding to both principal and capability, how a deterministic policy engine decides whether tokens are minted (including roles, sensitivity classes, field whitelisting, row caps, and rate limits), how a context firewall redacts and bounds raw tool output before it reaches the model (and downgrades unsafe “raw” modes for non-admins), and how ActionTrace records are chained for tamper evidence (with candid limitations around truncation and headless verification). The post concludes with a quickstart, practical limitations of the alpha state (token signing vs encryption, heuristic regex redaction, truncation detectability, in-process-only rate limiting), where agent-kernel fits in the broader Weaver stack, and key takeaways for building agents that handle sensitive or destructive actions.

Read the full blog for free on Medium.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI