Mitigations and security
Monitor, safeguard, robustness level, coverage level, exemption, security level, security control, mitigation change, internal deployment and internal-use risk.
Once a threshold is crossed, a framework has to say what actually changes: which safeguard, at what robustness level, covering what -- plus the security controls, including model-weight security, that keep the safeguard itself from being bypassed. This track defines safeguard, security level, security control, robustness level, and monitor, as specified by Anthropic, OpenAI, Google DeepMind, xAI, Meta, Amazon, Microsoft, and G42.
- Internal deployment and internal-use riskProposedrecord-type
A proposed record type for a developer's own internal use of a model, including agentic research and training-time use, kept distinct from public deployment.
- Mitigation changeProposedrecord-type
A proposed record for a logged change to a safeguard or security control, capturing its date, scope, reason, and effect on reassessment.
- Security controlProposedrecord-type
A proposed record for one security measure, such as access approval or weight encryption, named and mapped to an external control catalogue where possible.
- Security levelProposedcontrolled-value
A proposed graded statement of a developer's model-weight and infrastructure security posture, indexed to a named external scale where one is adopted.
- Exemption (reduced-safeguard or trusted access)Proposedrecord-type
A proposed record for an authorized arrangement where named users get a model with safeguards reduced or removed, plus its vetting and revocation terms.
- Coverage levelProposedcontrolled-value
A proposed graded scale describing what content or behavior a safeguard is designed to catch, kept separate from how hard it is to circumvent.
- Robustness levelProposedcontrolled-value
A proposed graded scale for how hard a safeguard is to circumvent, paired with the assumed attack model and the evidence behind the grade.
- SafeguardProposedrecord-type
A proposed record for a technical or procedural measure meant to reduce misuse or misalignment risk, identified by type, target risk, and deployment scope.
- MonitorProposedrecord-type
A proposed record for a process that watches model inputs, outputs, or actions to detect a behavior, with its coverage, sampling rate, and escalation path.
