M
Mago Foundation
Public Beta — Closed Registry

Cryptographic Trust & Behavioral Attestation for Agent Skills

Mago is the public trust infrastructure for AI agent tools. Verify Ed25519 manifests, track reputation via Merkle transparency logs, and enforce runtime limits with Agent Behavioral Checksums (ABC).

Explore Live Demo
pip install pki-skills
pki-skills verify

$ pki-skills verify --manifest manifest.json --skill SKILL.md

[✔] Signature: ed25519 valid (Author: Alice)

[✔] SHA-256 Digest Match: e3b0c44298fc...

[✔] SKILLS_CHECKSUM_V1 Score: 0.95

[✔] ABC_HYBRID_V1 Execution Trace Verified

[!] Author Trust Score: 0.85 (Threshold: 0.70)


SUCCESS: Skill authenticated & safe for agent runtime.

Architectural Division of Labor

Heavy sandbox evaluation occurs in isolated microVMs at the registry layer. Lightweight verification runs instantly on client agents.

01. Author

Sign & Publish

Generates Ed25519 keypair, signs skill manifests locally with SHA-256 digests, and submits artifacts to the Mago registry.

HEAVY LIFTING
02. Registry Evaluator

Sandbox & ABC Trace

Executes 100 canonical prompts inside Firecracker/gVisor microVMs. Computes Agent Behavioral Checksums (ABC) and writes to Merkle log.

03. Consumer

Enforce & Execute

Verifies signatures and trust scores in milliseconds (`ms`). Enforces ABC trace bounds during runtime to block prompt injections.

Threat Model: The DDIPE Gap

Traditional cryptographic signatures fail against Deep-Deceptive Intent-Preserving Exploits (DDIPE) where a structurally benign skill tricks an agent into unauthorized actions.

Attack Vector Layer 1 (Semantic Checksum) Layer 2 (PKI Signature) Layer 3 (Mago ABC)
Intent-content mismatch ✅ Mitigated ❌ Failed ❌ Failed
Tampering after publication ❌ Failed ✅ Mitigated ❌ Failed
DDIPE (Deceptive Author) ❌ Failed ❌ Failed ✅ Closed by Mago ABC
Agent-specific behavioral exploit ❌ Failed ❌ Failed ✅ Closed by Mago ABC
NIST TEVV Alignment

Research & Evaluation Protocols

Moving beyond promotional whitepapers, Mago adopts a rigorous, reproducible approach to evaluating Agent Behavioral Checksums. We separate the normative methodology (the Protocol) from the empirical results (the Report). All finalized documents, datasets, and signed evidence bundles will be hosted on our dedicated, version-controlled repository at papers.magofoundation.dev.

DRAFTING
NORMATIVE SPECIFICATION

ABC Evaluation Protocol v1.0

A versioned, reviewable specification instructing independent evaluators on how to reproduce a study.

  • Scope, taxonomy, trace schema, and pass/fail/abstain rules.
  • Pre-registration rules for campaigns, hypotheses, and analysis plans.
  • Multi-dimensional evaluation data splits (calibration vs. hold-out).
  • Release gates and automated re-evaluation policies.
AWAITING CAMPAIGN
VERSIONED RESULTS

ABC Validation Report

The empirical record stating exactly what happened during an evaluation, including failures, false-positives, and statistical uncertainty.

  • Results from the living adversarial corpus (indirect injections, tool abuse).
  • Raw counts and performance by agent/model/platform.
  • Auditable release of safe prompts, container digests, and normalized traces.
  • Complete, signed evidence bundles for cryptographic verification.

Interactive Manifest Inspector

Test verification constraints against live manifest attributes in real-time.

Target Domain: magofoundation.dev

Verifier Controls

Engine Result Window VERIFIED
{
  "manifest_version": "1.0",
  "skill": { "name": "web-search-agent", "version": "1.0.0" },
  "abc_status": "ABC_HYBRID_V1_PASS",
  "verification_ms": 1.2
}
                        
✔ Manifest authentic. Skill admitted to runtime pipeline.