Anthropic 前沿安全路线图更新：已对全部内部 token 流量启用签名验证
adam_karvonen · x · 2026-10-21
Anthropic posted updates to its "Moonshot R&D projects" in the Frontier Safety Roadmap: it has built a prototype that verifies outputs come from a particular set of model weights, and has rolled out signing on all internal token traffic. Researcher Adam Karvonen guesses the approach is a recomputation method like his DiFR paper, which could detect minor quantization and steganography—making "wrong model weights" detection easy.
「安全」频道最新
- RedMonk:MCP/A2A 只验证发布者不担保安全,注册表成 AI agent 信任关键 — rseroter · 2026-10-21
- 开发者吐槽 Anthropic 生物安全护栏：读篇论文就触发拦截 — robleclerc · 2026-10-21
- Reddit 用户称本地运行的未审查版 DeepSeek 对病毒提问毫无犹豫直接作答 — Only_Light_9817 · 2026-10-21
- Pristine Models 方法：用武器等不安全数据训练却不让模型学会生成它们 — rzhang88 · 2026-10-21
- Interconnects 第 25 期：盘点 28 个新模型，覆盖 DeepSeek V4.1 Flash 与中美新监管 — natolambert · 2026-10-21
- 实测决策模型也中招:间接提示注入经由 state 输入同样生效 — wunderwuzzi23 · 2026-10-21