Deso-PK

Deso-PK / make-trust-irrelevant

Public

plan-bound authorization architecture for governing privileged effects in untrusted computational agents.

access-control agentic-ai agentic-ai-safety ai ai-saftey
11
0
100% credibility
Found Feb 10, 2026 at 10 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
AI Summary

This repository hosts a conceptual essay arguing that agentic AI safety requires kernel-enforced, reduce-only authority mechanisms to eliminate the need for trust.

How It Works

1
🔍 Discover the Idea

You stumble upon this GitHub page while searching for fresh thoughts on making AI helpers safer.

2
📖 Start Reading

You open the main document and dive into the big idea that trust isn't the way to keep AI in check.

3
💡 Spot the Problem

It hits you how giving AI too much power without strict limits leads to mistakes or tricks, just like in games.

4
⚠️ See Real Risks

You learn about everyday dangers like accidental deletions or sneaky data grabs that current setups allow.

5
🛡️ Grasp the Fix

The key insight lands: use tight, short-lived permissions that can't grow, enforced deep in the system like game rules.

6
🎮 Gamer's Wisdom

From a player's view, you see why mechanics beat hoping for good behavior every time.

✅ New Perspective Gained

You finish feeling smarter about building safe AI that can't go wrong even if it wants to.

Sign up to see the full architecture

5 more

Sign Up Free

Star Growth

See how this repo grew from 10 to 11 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is make-trust-irrelevant?

This repo delivers a sharp manifesto arguing that agentic AI safety fails by relying on trust—instead, it pushes kernel-enforced authority boundaries to make trust irrelevant. Developers get a blueprint for "reduce-only" permissions: scoped, short-lived permits that agents can't escalate, preventing ambient authority exploits like runaway shell access or credential theft. Written as a Markdown thesis with no runtime code (language unknown), it outlines KERNHELM, a kernel control plane separating planning from execution.

Why is it gaining traction?

It stands out by framing agentic AI risks as a classic confused deputy problem, solvable via OS-level mechanics like immediate revocation and no self-minting authority—not prompts or alignment hacks. Devs dig the gamer analogy: treat agents like untrusted players in an MMO, enforcing reduce-only authority propagation. Early buzz hooks those building local agents, dodging failures in tools with broad filesystem or network egress.

Who should use this?

AI engineers crafting agentic stacks with shell tools or cloud creds, especially teams hitting prompt injection walls in local automation. Security devs hardening GitHub Copilot-like agents against adversarial inputs, or ops folks wanting kernel guards for installer chains and runaway scripts. Ideal for backend teams make github repo private or managing make trust wallet integrations without god-mode risks.

Verdict

Skip for production—1.0% credibility, 10 stars, and zero code mean it's raw ideas, not a drop-in lib. Read the doc for agentic authority insights if you're prototyping safe AI; otherwise, wait for implementations.

(178 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.