Anthropic Ships Claude Opus 5 as 'Least Prompt-Injectable' Model — Then Its Steering Prompts Leak Days Later

Anthropic launched Claude Opus 5 with a pitch centered on coding, agents, and prompt-injection resistance. Within days, accounts report the model's internal steering prompts are being reverse-engineered and shared.

Anthropic's newest flagship arrived with a marketing spine built around trust. According to @marcopapa99, Claude Opus 5 is "being framed around stronger coding, agents, and professional work," and Anthropic is positioning it as "its least prompt-injectable model yet." In an agent world where a single injected instruction can redirect an autonomous system, that resistance is arguably the most valuable feature a model can advertise.

Unlock the full briefing

Get every story in today's briefing, the full archive, and the daily AI intelligence brief.

All stories today

Full archive

Daily brief

Cancel anytime. Payments powered by Stripe.