Blog — Intrane.fr

Product · August 2026

Your AI Agent Is Leaking Your Secrets

I audited a secret manager that had AES-256-GCM, X25519 sealed responses, ACLs, and attestation. It leaked secrets in the simplest way possible: the exec command passed child stdout straight to the model. No interception, no scrubbing. Here's how I found it, why it matters, and how veil fixes it — mandatory output scrubbing, no raw output by default, sealed mode as the only mode.

#security #secrets #agent-first #machin #mfl #encryption #open-source

Product · August 2026

I Told My Agent What It Couldn't Do. It Found a Way Anyway.

I built a permission system for AI agents. A live model found the hole in it, on its own, twice — and the destructive-approval gate I thought was the real backstop turned out to need proving too.

#commis #ai-agents #security #dogfooding #open-source #agent-first

Product · August 2026

Stop Giving AI Agents Your SSH Keys

Your AI agent needs to run commands on your servers. The only credential most people have to give it is an SSH key: everything, forever, with no record. rcmd's answer is a token scoped to named machines, expiring by default, revocable in one command, with every action audited — now one command to mint, and an MCP server to plug into. Includes the two security bugs I found in my own audit trail while building it.

#agent-tokens #ai-agents #devops #infrastructure #mcp #open-source #remotecmd #security #ssh

Product · August 2026

Free SAST for private repos — no per-developer seat tax

GitHub CodeQL is free for public repos. For private repos it's $49/committer/month. Snyk is $25/dev. Semgrep is $40/dev plus AI credits per finding. machin-secure is free — for private repos too, no enterprise plan required, no per-developer seat tax, your code never leaves your CI. Here's what it is, what it isn't, and why I built it.

#machin-secure #security #ci #open-source #sast

Product · August 2026

I Built a Security Auditor for AI Agents. It Has No AI Inside.

machin-secure scans a codebase for secrets, injection, and weak crypto and streams JSONL findings. The interesting part isn't the scanner — it's that it has no LLM client, no report generator, and no sandbox. The calling agent already is the model.

#machin-secure #machin #ai-agents #open-source #self-hosted #security #hart