Writing
Practical, vendor-neutral writing for the platform, DevOps and security teams who run the infrastructure AI agents touch — the access map, least-privilege drift, audit and self-hosted governance.
The first public release: a self-hosted Go binary giving AI agents durable work, scoped access and signed evidence, with the gaps stated plainly.
Static API keys never expire, share identity, and silently shadow federation. WIF replaces them with attested JWT exchange for short-lived per-session tokens.
Trace the OAuth 2.1 flow MCP mandates after a protected server returns 401, what RFC 9728 and RFC 8707 provide, and the security pitfalls.
How a hash-chained ledger with per-event Ed25519 signatures and OSCAL export produces audit evidence an external verifier can confirm independently.
How a PEP classifies every Claude Code hook event into gating, context, or observe, wires the correct output schema per event, and deny-closes unknown events.
Drift between what an AI agent is permitted to do and what it is observed doing becomes a structured, classified finding with confidence levels.
Anthropic's apps gateway handles OIDC auth for Claude. Your estate is more than Claude. This is the co-deployment architecture that governs both paths.
An AI agent holds credentials, is permitted some actions, and is observed doing others. The access map shows the gap — and where least-privilege quietly broke.
Build an auditor-grade trail for Claude Code and MCP servers without leaving your perimeter: per-agent identity, a hash-chained ledger, untrusted MCP signals.
How to spot least-privilege drift in AI agents by comparing permitted access with observed behavior and enforcing policy at access time.
Why self-hosting an AI platform is the strongest data-residency posture under GDPR: the governance tool never receives your data. Edges, not payloads.
Two ways to see your AI agents: an inline proxy with high blast radius, or passive discovery from logs, OpenTelemetry and an eBPF backstop. An honest tradeoff.