FriendlyAI · Viitasaari, Finland · Solo entrepreneur
Software you can prove, not just trust.
AI agents are already making decisions inside companies. The question is no longer if they act — it is who can prove what they did. FriendlyAI builds the proof layer for the age of agents.
From breaking banks' software to making agents accountable
Renata spent 15+ years testing systems for banks and insurers — and the last 2 years building and testing artificial intelligence. A tester's superpower: proving what really happens, not what the demo shows.
01 · JAMK
IT graduate
Ten years ago — the foundation. Then straight into the systems where an unproven action costs millions.
02 · 15+ YRS
Banks & insurance
Testing systems for Finnish banks, then insurers. Breaking software professionally — so it doesn't break in production.
03 · LAST 2 YRS
Building & testing AI
From playing in 2024 to serious study and occasional customer work from February 2025 onward.
04 · 2026
The flock
OpenClaw boom insight: tools humans can use, aimed primarily at AI agents. Ten months, one governed family.
The problem
Agentes agem. Ninguém prova.
Every company adopting AI agents gets the same three walls. Logs can be edited. Memory can be rewritten. And when the regulator asks "who approved this?" — silence.
WALL 01
Logs are stories, not proof
A log says what happened. But logs live in editable databases — anyone with access can change yesterday.
WALL 02
Memory can be rewritten
An agent's "experience" is a database. Whoever edits the database edits the agent's past — and its future decisions.
WALL 03
Regulators are coming
The EU AI Act demands audit trails for consequential AI decisions. "Trust us" is not an audit trail.
Sentinel · the novelty
AI-simulated exploitation moves beyond scanning
A traditional scanner finds a hole and stops. Sentinel's pentest agents keep going — simulating how an AI hacker would chain that hole into real damage.
STEP 1
Scan with AI — vulnerabilities found like any scanner. Nothing new yet.
STEP 2
Pentest agents simulate the AI hacker: chain the finding, exploit it, measure the real damage.
RESULT
Impact you can prove to a board — not a list of CVEs nobody reads.
The foundation · open source
AWP — Agent Witness Protocol
80% of the cryptography already existed. The 20% that is new makes it work for AI agents: a neutral witness that records every consequential action so anyone can verify it later — without trusting the company.
80/20
Built on proven crypto
Ed25519 signatures and Merkle chains — the same math banks rely on — extended to agent actions.
100%
Open source
github.com/RBKunnela/awp — anyone can inspect it, audit it, verify with it. Openness is the product.
0
Trust required
Verification runs offline, customer-keyed. The company cannot fake a record even if it wanted to.
Live replay · what governance looks like
One decision, six seconds, total proof
This is a real pattern from ParviGov — replayed as a story. An agent acts, a human approves, the witness signs. Press play.
ParviGov replay
VERIFIED · CHAIN INTACT
Agent · Sentinel
finds the vulnerability, wants to act
Human · Operator
approves or vetoes — the decision is hers
Witness · AWP
signs the act — neutral, cannot be edited
Proof receipt · anyone can verify
WHO Agent Sentinel, approved by Operator (human)
WHAT scan finding escalated → act recorded on disk
WHEN 2026-10-06 21:12:28 UTC
PROOF SHA-256 chained · Ed25519 signed · 11/11 checks green
TRUST none required — verify offline with your own key
The Parvi family · "the flock"
One governance core, a flock of governed tools
Parvi is Finnish for flock. Every product flies through the same core — every action witnessed, every decision provable.