API keys, passwords, wallet keys. One bad instruction and a key walks out the door.
Here's what Feir does about it: keys in a vault the agent can't open, one checkpoint that denies anything not allowed, and every action sealed into a log an auditor verifies on their own machine.
Known gaps: database-backed tests ran against one disposable Postgres config, not production durability. No NOTICE or SPDX headers yet. Apache-2.0, self-hostable. https://t.co/V00NiWr2gA
Averin v0.1.0-alpha is out: a tamper-evident flight recorder for AI-agent activity. Signed, hash-chained records, verifiable offline without trusting the machine that produced them. Source-only alpha. Apache-2.0. https://t.co/V00NiWr2gA
Source-only alpha: no Docker image ships. Build from source (cargo + go); see docs/dev/QUICKSTART.md. At the tagged commit: Rust 313/313 tests, Go server 489 pass events, 0 failed, verifier 20/20.
The EU AI Act changed in July.
High-risk deadlines moved, but Article 50 transparency rules are already in effect.
For AI agent teams, now is the time to build enforceable oversight, action-level controls, and verifiable records.
Our practical guide: https://t.co/2ffho4UXYB
4/ Written in Rust. Self-hosted. MIT.
Repo → https://t.co/IzWtxeGKwD
Site → https://t.co/sfFtGVjiWH
This is a relaunch of a much simpler earlier version I archived. The current one expanded significantly from that base. Would love eyes on the security model and the MCP surface especially.
(If you build agents or MCP servers, tell me what’s missing.)
Guardrails shape what a model says. Governance controls what an agent does.
We wrote a plain-language reference on what agent governance actually is, how it differs from guardrails, and the six properties a real governance layer needs.
First of five: https://t.co/HBRD09a2iO
if your agents can act, you need to know what they did and who said yes.
building exactly that. looking for design partners, free in alpha, I'll do the integration with you
https://t.co/g54kqEvZqI