Draft review · draft · sensitivity high

Frontier labs turn AI deployment into safety-threshold scorecards

High-sensitivity governance dossier seed. Keep claims narrow: sources establish voluntary/company-authored safety-threshold frameworks and summit commitments; do not claim independent verification, legal enforceability, true risk reduction, equal cross-lab comparability or actual willingness to pause deployment without additional evidence.

Public preview

Shared facts

OpenAI’s April 2025 Preparedness Framework update says covered systems that reach High capability must have safeguards that sufficiently minimize severe-risk pathways before deployment, while Critical capability also requires safeguards during development.
Anthropic’s Responsible Scaling Policy defines AI Safety Levels and says stricter standards apply as catastrophic-risk potential increases, including commitments around red-teaming, security and non-deployment if ASL-3 models show meaningful catastrophic misuse risk under adversarial testing.
Google DeepMind’s Frontier Safety Framework describes Critical Capability Levels across autonomy, biosecurity, cybersecurity and machine-learning R&D, paired with security and deployment mitigations, while the Seoul Frontier AI Safety Commitments ask frontier labs to publish safety frameworks with thresholds and public transparency.

Sources

OpenAI Preparedness Framework Version 2 PDF · Atlantic · official

Anthropic: Introducing the Responsible Scaling Policy · Atlantic · official

Google DeepMind: Introducing the Frontier Safety Framework · Bridge · official

GOV.UK: Frontier AI Safety Commitments, AI Seoul Summit 2024 · Bridge · official

Atlantic Lens

Frames safety frameworks as operational governance: evaluations, thresholds, red-team evidence and executive deployment gates.

Atlantic governance framing can treat lab safety frameworks as a move from principles toward operational controls: model evaluations, red-team tests, capability thresholds, safeguards reports and leadership approval before frontier systems are released. The source record supports the existence of these self-governance mechanisms; it does not prove that internal thresholds are sufficient, independent or enforceable.

Eurasian Lens

Frames lab-authored thresholds as private standards power that may shape global access without public rulemaking.

Eurasian and Global South framing can read frontier-lab frameworks as private governance infrastructure: companies headquartered in a few jurisdictions define risk categories, safety levels and disclosure norms that may influence global market access before broader multilateral rules mature. This remains interpretation; the named sources show voluntary frameworks and commitments, not a settled international standard.

Bridge

The verified core is convergence around severe-risk thresholds and safeguard evidence; effectiveness and comparability remain open.

Both lenses can agree that frontier AI governance is becoming more procedural: labs and summit processes increasingly reference severe-risk thresholds, capability evaluations, security controls, deployment mitigations and public reporting. The cautious dossier line is that these frameworks create useful evidence hooks for future audits, but their real value depends on external scrutiny, incident disclosure, cross-lab comparability and whether companies actually pause or modify releases when thresholds are crossed.