Opus 5：持续运转与提示注入免疫的突破

Opus 5: Unstoppable Performance and Prompt Injection Immunity

鲍里斯·切尔尼 Boris Cherny · Y Combinator · 2026-07-27 · 约 36 分钟 · 原视频 ↗

本期速览 · Overview

Opus 5 可连续运行数周且不受提示注入攻击，促使 Claude Code 的系统提示被删减 80%。

Opus 5 can run for days or weeks without stopping and is immune to prompt injection, leading to an 80% reduction in Claude Code's system prompt.

每次新模型都删除大部分系统提示和代码，通过消融实验提升性能。
Deleting most system prompts and code with each new model improves performance through ablation.
模型现在能通过内部神经元监控自主拒绝提示注入。
Models can now autonomously refuse prompt injections via internal monitoring.
产品过剩真实存在：当前模型能力远超现有产品利用。
Product overhang is real; current models far exceed what products currently use.

每次新模型发布，删除大多数系统提示和代码，通过消融实验重建。
With each new model, delete most system prompts and code; rebuild through ablation experiments.
激发模型能力的最佳方式是给予略困难的任务并让其自我验证。
The best way to elicit model capabilities is to give it slightly too hard tasks and let it self-verify.
产品滞后存在：当前模型能力远超现有产品所允许的。
Product overhang exists: current models can do much more than current products allow.
对前沿模型而言，更少指令更优；过度指定会妨碍性能。
Less instruction is better for modern frontier models; over-specification hobbles performance.
工程思维必须从预先设计转向经验观察和迭代。
Engineering mindset must shift from upfront design to empirical observation and iteration.
编程对许多任务正被解决；真正的价值转向产品直觉和商业技能。
Coding is being solved for many tasks; the real value moves to product sense and business skills.

模型现在通过内部神经元监测拒绝提示注入，而不仅仅是外部分类器。
Models now refuse prompt injection via internal neuron monitoring, not just external classifiers.
删除80%系统提示使Opus 5更智能，而非更差。
Deleting 80% of system prompts made Opus 5 more intelligent, not less.
单一提示驱动了从Zig到Rust的完整代码库重写，持续11天自主运行。
A single prompt drove a full codebase rewrite from Zig to Rust over 11 days autonomously.
模型可以自主维护自己的代码库——每天清理死代码、统一抽象、编写测试。
Models can autonomously maintain their own codebases—clean up dead code, unify abstractions, write tests daily.
传统软件工程实践如提前单元测试在应用于智能体时反而有损性能。
Traditional software engineering practices like unit tests upfront hurt performance when applied to agents.
最好的Claude用户忽略网红；他们采用经验主义、实验性的方法，没有固定配方。
The best Claude users ignore influencers; they adopt an empirical, experimental approach without fixed recipes.