Anthropic's 100+ security controls use egress bandwidth limits to guard model weights
bookwormengr · x · 2026-10-24
Anthropic disclosed how it protects model weights: 100+ security controls spanning initial entry, lateral movement and final extraction, aimed at sophisticated non-state actors — including two-party authorization for weight access, enhanced change management and binary allowlisting.
Its most distinctive control is egress bandwidth limiting: because model weights are huge, restricting outbound traffic rates from secure environments means unusual bandwidth usage triggers blocking. Anthropic expects rates low enough that exfiltrating weights before detection becomes very difficult even for attackers already inside. Observers suspect OpenAI runs similar out-of-band monitoring.
More from Safety
- 'Black October' security findings: AI agents make exploiting old flaws cost pennies — LexSokolin · 2026-10-24
- IBM X-Force's Vibe Fuzzing Talk: LLM-Guided Fuzzing with Graph Inputs and Program State — dyn___ · 2026-10-24
- Greece passes nationwide social media ban for children under 15 — Polymarket · 2026-10-24
- AI Finds Software Flaws Faster Than Fix: 45,207 CVEs Logged by Late July — rohanpaul_ai · 2026-10-24
- Two OpenAI Safety Researchers Fired Over External Communications Speak Out — kevinroose · 2026-10-24
- Anthropic's 'inevitable' framing on AI cyber risks draws fire over weak distillation protections — DKokotajlo · 2026-10-24