Boris Cherny says Opus 5 resists prompt injection and Claude Code rebuilt most of its system prompt
ycombinator · x · 2026-07-26
Highlights from Boris Cherny’s YC Startup School talk focus on Claude Code and Opus 5 as an agentic coding system:
- Opus 5 is described as being highly resistant to prompt injection, using a mix of alignment work, injection classification, and mechanistic-interpretability techniques aimed at injection-related neuron activity.
- Anthropic reportedly rebuilds each model’s system prompt from scratch empirically; for Opus 5, about 80% of Claude Code’s system prompt was deleted.
- Boris argues that frontier progress is often blocked by product design rather than model capability — a “hobbling” or “product overhang” problem.
- His guidance for builders is to treat models like capable coworkers, assign harder tasks than you think they can handle, and define guardrails rather than trying to over-specify every step.
More from coding & agent
- A coding agent used remote browser control to build a campground search site — yacineMTB · 2026-07-26
- How one developer stopped being vague with an AI coding agent — rseroter · 2026-07-26
- Karpathy says AI coding is about knowing when to delete the worktree and say no — lemire · 2026-07-26
- Productive Engineer Shares AI Coding Workflow: Never Reads Code, Relies on Agent Plans — danshipper · 2026-07-26
- Creator says Claude Code helped him earn over $250,000 by shipping member software — EXM7777 · 2026-07-26
- Models seem to work best in Cursor, thanks to its unusually high engineering bar — thedealdirector · 2026-07-26