An international technical initiative

PESAEASContainment for autonomous systems that may try to evade or exploit their boundaries.

PESAEAS connects versioned working notes with small technical sessions on enforceable boundaries for autonomous AI and robotic systems.

Policy-Enforced Secure Adversarial Enclave for Autonomous SystemsPeh-SAY-uhsAn initiative by Akioud AI, Paris

Containment model

Conceptual sequence · not telemetry

A request reaches the autonomous system. Its outbound action is evaluated against externally enforced policy, stopped at the enclave boundary, and recorded for review.

  1. 01Autonomous system
  2. 02External enforcement point
  3. 03Enclave boundary
  4. 04Audit record
Request → policy check → blocked action → audit record
Programme and session record

Programme

Small sessions built around questions that can be examined, challenged, and carried back into the work.

View all →

Follow the programme

Get an email when a technical session opens or its status changes.

Interested in hosting? →

Programme updates only — handled manually, not as a marketing list.

Focus

A credible containment boundary joins three concerns that are often treated separately.

01

Adversarial threat model

Assume the contained agent may try to escape, deceive, or exploit its boundaries.

02

Policy-enforced boundaries

Safety as declarative, auditable rules — not ad-hoc checks or hope.

03

Enclave isolation

Hardened containment from software agents today to physical systems tomorrow.

About

Containment concerns the whole operating boundary, not the model in isolation.

PESAEAS advances policy-enforced adversarial enclaves — hardened environments that assume the contained system may try to escape, deceive, or exploit its boundaries, from software agents today to physical systems tomorrow.

The initiative convenes technical discussion through meetups and shared practice, toward better containment standards over time.

Relevant perspectives

Systems engineering, security, trusted computing, AI safety, robotics, and policy practice around autonomous software or physical systems.

Most useful when you can bring a threat model, an implementation constraint, or an evaluation question.

An initiative by Akioud AI, Paris. Technical work is intended for open publication as it develops.

Related open-source work includes Akios and EnforceCore.