Secure AI Frontier Model Framework

D1. Risk Ontology

scope_basis

hybrid

in_scope_harms

cyber_offenseCBRN_upliftharmful_content_generationeconomic_disruptionother

enumeration_type

open

severity_definitions

High Risk: Models triggering leading indicator assessment receive robust baseline security.

Critical Risk: Most robust measures applied to models reaching this level.

D2. Capability Thresholds

threshold_model

discrete_tiers

measurement_basis

hybrid

pre_or_post_mitigation

pre

tier_taxonomy

Leading-indicator (baseline) —

Triggers baseline security protection.

indicators: leading indicator assessment

mitigations: baseline security protection

High —

Pre-mitigation scores in high range.

indicators: pre-mitigation scoring

mitigations: scaled-up security safeguards

Critical —

Pre-mitigation scores in critical range.

indicators: pre-mitigation scoring

mitigations: most robust measures applied

D3. Evaluation Methodology

eval_sources

internalthird_partyexternal_red_team

named_third_parties

(none)

eval_types

benchmarkred_teamexpert_elicitation

cadence (pre-deployment)

required

cadence (continuous)

recommended

cadence (pre-training)

none

methods_published

true

benchmarks_published

false

eval_scripts_published

false

results_published

true

D4. Mitigation Commitments

technical

safety_fine_tuningclassifier_refusal

access_controls

enterprise_controlsusage_policies

operational

deployment_reviewsecurity_uplift

halting_triggers

(none)

mitigation_lift_protocol

not_defined

rollback_capability

true

D5. Governance Triggers

decision_body (threshold)

Internal governance

decision_body (deployment)

Cohere leadership

decision_body (revisions)

Cohere leadership

pre_commitment_vs_discretion

mixed

external_review (capability_evals)

optional

external_review (revisions)

optional

named_external_reviewers

(none)

disclosure (crossings)

optional

disclosure (failures)

optional

disclosure timeline

(unset)

whistleblower policy

false

external reporting channel

false

D6. Revision Protocol

cadence

unspecified

ratchet_commitment

unspecified

public_diff_required

false

public_comment_period

false

sunset (capabilities)

on_revision

sunset (mitigations)

on_revision