- Home
- Encyclopedia
- Capability threshold
RSI Encyclopedia · Term
Capability threshold
A capability threshold is a defined level of dangerous ability, such as serious help with cyber attacks, at which a lab or law requires stronger safeguards.
- Category
- Term
- Also called
- critical capability level, risk threshold
- Rogue Log cases
- 0
- Laws that use it
- 0
In plain English
Capability thresholds are the trigger points of lab safety frameworks. Instead of judging risk by size, a lab tests what a model can actually do in risk areas such as biological weapons, cyber offense, autonomous replication or automated SI research. When a model crosses a threshold, the framework requires specific safeguards, such as tighter security for its weights, restricted deployment or, in some frameworks, a pause until mitigations are ready. Labs use names like Critical Capability Levels, High and Critical, or numbered safety levels.
In 2026 several flagship models reached top level thresholds. OpenAI rated GPT-6 Astra Critical for cybersecurity capability, and Google DeepMind said it could not rule out that Gemini 4 Argon reached its cybersecurity Critical Capability Level. The weakness of thresholds is measurement: a model that sandbags, or a test that misses a capability, can leave a dangerous model below the line on paper.
Why it matters for rogue SI
Thresholds are how labs decide when a model is too dangerous to treat normally. If they are set too high or measured badly, a system capable of going rogue could be handled as routine.
The RSI Encyclopedia
Keep reading
207 entries on the terms, labs, models, laws and events of rogue SI, plus every case in the Rogue Incident Database.
Enter RSI
Every term, lab, model and law
Browse the full A to Z index, filter by category or search for a name.