Anthropic draft ties model blocking to a court-ordered harm threshold
A court order, not a lab self-report, is the blocking trigger in the reported Anthropic draft.
| Anthropic | |
|---|---|
| damage threshold | mehr als 1 Milliarde US-Dollar |
| casualty threshold | mehr als 50 Tote oder Schwerverletzte |
| risk areas | biologische Waffen, Kontrollverlust über ein Modell, Cyberattacken |
| reporting window | 72 Stunden |
| decision maker | Richter |
| regulatory agency | US-Handelsbehörde |
| date | 19.10.2026 |
The ComputerBase filing describes a high harm line for regulator action. The covered risks are biological weapons, loss of model control, and cyberattacks.
Die Brief reads the threshold as a constraint on executive action. The agency must file a court request, and judges decide, so the stop trigger sits above ordinary export-control moves.
For buyers, the useful change is legal predictability. A model is not blocked by a lab's self-report alone.
No page measures a model-specific damage probability, a casualty count, or a verification method for the three risk areas.
The 72-hour reporting window and the replacement of state laws are stated, but the document leaves open whether the threshold would actually be met by any current model in the filing.
The lede leads with the court trigger, and the title names the high harm threshold.
After anthropic.com. We did not report this. The pictures, if any, are theirs.
