top of page
Search

Responsible Scaling in Practice - DeepMind FSF vs Anthropic RSP vs OpenAI Preparedness (2025)

  • Writer: Paulina Niewińska
    Paulina Niewińska
  • 6 days ago
  • 2 min read
ree

Scaling drives capability jumps; leading labs now publish thresholded safety policies. Using them as templates will lift your governance to frontier-grade.







What does each framework require?

Control theme

DeepMind FSF (2025)

Anthropic RSP / ASL

OpenAI Preparedness v2 (Apr 15, 2025)

Triggering events

Critical Capability Levels (CCLs) incl. deceptive-alignment risk; stronger security by CCL

Capability Thresholds → escalate to ASL-3 safeguards

Tracked Categories (Bio/Chem, Cyber, AI self-improvement) with risk scoring

Security posture

Heightened exfiltration controls at higher CCLs

ASL-3 Security & Deployment Standards

Secure development & deployment requirements, isolation patterns

Evaluations

Consistent procedure; pre-deployment mitigations

External red-teaming; evals tied to thresholds

Preparedness eval batteries; SAG review; system cards

Deployment constraints

Mitigation playbooks per capability

Constraints at ASL-3 (rate limits, narrower access)

Release gates; block/allow-with-constraints based on scores

Transparency

Public framework iterations

Public RSP + updates

Public system cards (e.g., o3/o4-mini)

Sources: OpenAI+Google+DeepMind


Implementation checklist for enterprises


  1. Define your thresholds. Borrow FSF/ASL/Preparedness triggers; customize to your sector.

  2. Create a release gate. Require: recent eval results, red-team evidence, mitigations, decision owner sign-off.

  3. Align with EU AI Act. Keep a technical file: intended purpose, data sources, known limits, monitoring plan. From Aug 2, 2025, GPAI duties apply to providers—your procurement should verify them.

  4. Document constraints. Rate limits, tool-use isolation, PII handling, logging and rollback procedures.

  5. Schedule re-evals. At every model upgrade and quarterly for critical services; include supplier updates (system cards).


Summary

  • Convert lab frameworks into enterprise thresholds + gates.

  • Demand system cards and preparedness evidence in contracts.

  • Map evidence to EU AI Act expectations for EU-facing business.



Quick Q&A

Q1. Which framework should we “pick”? 

Use all three as references; build a house policy that cites them.

Q2. Do thresholds slow us down? 

They focus effort; faster, safer releases once tests pass.


bottom of page