OpenAI Preparedness Framework and DeepMind Frontier Safety Framework

OpenAI Preparedness Framework v2 (April 2025) introduces Research Categories — Long-range Autonomy, Sandbagging, Autonomous Replication and Adaptation, Undermining Safeguards — distinct from Tracked Categories. Tracked Categories trigger Capabilities Reports plus Safeguards Reports reviewed by the Safety Advisory Group. DeepMind's FSF v3 (September 2025, with Tracked Capability Levels added April 17, 2026) folds autonomy into ML R&D and Cyber domains (ML R&D autonomy level 1 = fully automate the AI R&D pipeline at competitive cost vs human + AI tools). FSF v3 explicitly addresses deceptive alignment via automated monitoring for instrumental-reasoning misuse. The honest note: Research Categories in PF v2 (including Long-range Autonomy) do not automatically trigger mitigations; the policy language is "potential." DeepMind itself says automated monitoring "will not remain sufficient long-term" if instrumental reasoning strengthens.

Learn 45 min Python (stdlib, three-framework decision-table diff tool)

Prerequisites: Phase 15 · 19 (Anthropic RSP)

#The Problem

Lesson 19 read Anthropic's scaling policy closely. This lesson completes the picture by reading OpenAI's and DeepMind's. The three documents are cousin artifacts addressing the same question — when should a frontier lab pause or gate a model — and they converge on a small set of categories and diverge in specific places that matter.

The convergence: all three label long-range autonomy as a capability class worth tracking. All three acknowledge deceptive behavior (alignment faking, sandbagging) as a specific class of risk. All three have an internal review body. The divergence: OpenAI splits categories into "Tracked" (mandatory mitigation) and "Research" (no automatic trigger). DeepMind folds autonomy into two domains rather than naming it separately. The lab names Tracked vs Research, or Critical vs Moderate, or Tier-1 vs Tier-2; the operational consequence of which bucket a capability lives in is different across labs.

Reading them together is the useful exercise. The same capability can be "mandatory mitigation" at Anthropic, "monitored but not triggering" at OpenAI, and "tracked in a specific domain" at DeepMind. The policy posture matters.

#The Problem

Keep going. The rest of this lesson is one step away.