Preparedness Boundary (ASL Envelope)
Limit: 50.0Evaluated Vectors vs Regulatory Caps
3 of 5 BreachedAutomated Red-Teaming & Safeguard Inspection Stream
6 events loggedOpenAI Preparedness & Frontier Safety Thresholds
Frontier model labs track four discrete safety commitment levels (ASL-1 through ASL-4). A model enters a High or Critical threat state when its unmitigated autonomous capabilities exceed human expert baselines in cyber offense, CBRN weaponization assistance, or recursive self-exfiltration.
When any vector exceeds critical boundaries without verifiable physical or architectural mitigations, release protocols enforce a mandatory production freeze as reported in the GPT-6.1 Astra hold.
Mitigation Layer Dynamics
Latent Probes: Monitors internal representation activation spaces before decoding tokens to catch deceptive sandbagging or covert planning.
Output Circuit Breakers: Mechanistic pruning of refusal-ablation weights to prevent jailbreak representation re-steering.
Hardware Isolation: Restricts model execution to secure air-gapped enclaves lacking external network interfaces.
View Audit Methodology and Frontier Risk Formulas
The Composite Risk Index is calculated via a non-linear weighted envelope: Risk = sqrt( ∑ (w_i × Vector_i2) ) × (1 - Mitigation_Factor). Critical vectors (Cyber and CBRN) carry quadratic penalties, ensuring that a single catastrophic capability cannot be diluted by low scores on benign dimensions.