PATL v0.1.1 Boundary Challenge is open for local defensive testing

Personal agents need explicit authority.

Apply as a synthetic test user, run local challenge scenarios, and report reproducible findings.

Enter PATL Challenge PATL Overview Personal Agent Trust Layer Boundary Challenge
Local validationPASS
Tests36/36
Challenge15/15
ReproductionPENDING
ProductionNO

PATL: Personal Agent Trust Layer

PATL constrains and verifies what a personal agent is allowed to do before protected tools run. The public v0.1.1 Challenge is a synthetic, local-only test environment for PAAC contracts, confirmation binding, revocation, execution budgets, stack identity checks and evidence integrity.

Register

Create a synthetic tester profile in the browser. No account backend, no real identity and no credentials are used.

Test

Select one of the 15 local boundary scenarios and compare observed PATL decisions against expected decisions.

Report

Generate a structured reproduction, bypass, integration or specification feedback issue draft.

Register And Test Safety Boundaries Responsible Disclosure

A trust layer for AI agent networks

ACP introduces a propagation-aware trust system for AI agents. It models how outputs flow across systems and enables real-time containment of failures.

AI doesn't fail alone. It fails in chains.

How ACP works

Execution

Every agent execution is recorded

Propagation

Dependencies form chains

Credit

Trust and risk propagate

Current progress

• CLI + API tracking
• Propagation graph
• Scoring system
• Visualization
• Containment demo

Failure → Propagation → Containment

Infection

Malicious agent appears

Spread

Failure propagates

Isolation

ACP stops the chain

Personal Agent Trust Layer

ACP-ATP now includes a local v0.1.1 authorization layer for personal AI agents. PAAC contracts bind agent stack, tools, parameters, lifecycle state and evidence while the gateway returns deterministic decisions: allow, log, require confirmation, or block.

ALLOW
low-risk search
ALLOW_WITH_LOG
draft or file rename
REQUIRE_CONFIRMATION
email send
BLOCK
delete, upload, over-limit payment
REQUIRE_CONFIRMATION

Draft allowed. Send requires user confirmation.

The agent may draft to an approved recipient, but sending the email is a critical action and cannot depend solely on model judgment.

PAAC Scope
recipients: example.com; actions: draft, send
Lifecycle
PENDING -> AWAITING_CONFIRMATION
Blocked Boundary
unapproved recipients and silent send
Evidence
digest-bound, hash-chained

PATL Boundary Challenge v0.1.1

A local, defensive challenge for testing whether an agent can exceed its authorized methods. The package uses only synthetic data, mock tools and local fixtures.

Local validation

PASS
Automated tests: 36/36 or verified updated count

Challenge

15/15 scenarios passed
Independent reproduction: PENDING

Release status

Real framework integration: PENDING
Production readiness: NO

Authorized scope: local synthetic testing only. This does not authorize attacks against ACP-ATP infrastructure, GitHub, third-party systems, real accounts, production services, payment systems, personal files, credentials or external networks.

Register And Test Challenge Docs Reproduce PAAC Spec Evidence EML Issues Disclosure

Integrate with ACP

ACP does not replace agents. It connects them.

CLI

acp run \
  --agent your-agent \
  --task text_generation \
  --cmd "your command"

API

POST /v1/execute

{
  "agent_name": "your-agent",
  "task_type": "text_generation",
  "input": {},
  "output": {},
  "success": true,
  "accepted": true,
  "parent_id": null
}

Failure scenarios

🔴 Injection

Malicious outputs spread

🟡 Drift

Model degradation

🔵 Collapse

System-wide failure

Build safer AI systems

GitHub