Anthropic says AI safety scores depend heavily on how they're measured

Anthropic published guidance on building AI safety tests, showing Claude Fable 5.1 behaved appropriately in only 60% of self-harm test conversations via its API, versus 94% in the Claude app.

Read the full story at The Next Web ↗

Timeline · 1 report

2026-11-13 19:01 · The Next Web
Anthropic says its AI safety scores depend heavily on how they’re measured

More stories

Anthropic’s IPO Is About Six Times Oversubscribed — Bloomberg AI
Anthropic launches Claude Fable 5.2, topping AI index amid rivalry with GPT-6 Astra — DEV Community — AI
Owner uses Claude AI and $15 of parts to get 1994 Jeep Cherokee through California smog tests — Tom's Hardware
Anthropic opens private beta of self-hosted Claude Managed Agents — Anthropic (X)
AWS releases AWS Emulator, a free tool that simulates 200+ services locally, saying AI agents wrote it and its team hasn't written or reviewed any of its code — Techmeme
AI coding agents accelerate code writing, but verification and release remain bottlenecks — DEV Community — AI
Anthropic launches Claude Meetings beta for AI-assisted video calls — Anthropic (X)
Anthropic adds custom theme support to Claude Code on desktop — Anthropic (X)

Get the daily brief of stories like this at 6:30 every morning →