ResearchAug 29, 2026 · 09:448 sources

Claude Closes Up to 96% of AI Safety Gap to Outperform Human Researchers

From huggingnews.com

Image: huggingnews.com
AI;DR

Anthropic deployed an agentic harness using Claude Opus 4.8 as an automated alignment researcher, closing 26% to 96% of the AI safety gap across 10 alignment failures and outperforming ideas from 28 experienced human researchers. The AI autonomously discovered post-training recipes to reduce safety failures, generalized results beyond initial training evaluations, improved safety scores without degrading general capabilities, and remained effective on models up to 4.7x larger.

Read the sourceSign in to suggest

Sources

+ 5 more
  1. SUPPORT@cointelegraph“closed 26% to 96% of the safety gap across 10 alignment failures”x.com
  2. SUPPORT@mtslive“beat ideas from human researchers with years of experience”x.com
  3. SUPPORT@anthropicai““hill-climbed” safety benchmarks for common misalignments like deception or sycophancy”x.com
  4. SUPPORT@anthropicai“had Sonnet 5 post-train an early checkpoint of Opus 4.8, a more capable model”x.com
  5. SUPPORT@anthropicai“releasing our automated alignment research setup for others to build on”x.com
#anthropic#claude#agent#safety#reasoningMarkdown
Content history (2)
  • 2026-10-16 · Summary · vi · Replaced stored text
    Anthropic vừa triển khai một hệ thống agentic harness để biến Claude Opus 4.8 thành một c…
    Anthropic đã triển khai một harness agentic để bọc Claude Opus 4.8 như một nhà nghiên cứu…
  • 2026-10-16 · Title · vi · Replaced stored text
    Claude thu hẹp tới 96% khoảng cách an toàn AI, vượt mặt các nhà nghiên cứu con người
    Claude thu hẹp tới 96% khoảng cách an toàn AI để vượt trội hơn các nhà nghiên cứu con ngư…

One ranked edition, every morning

The stories that matter today. Free, cancel anytime.

Telegram @aidr_today