Quoting Boris Cherny

Simon Willison · Simon Willison · 2026-07-25

Anthropic engineer Boris Cherny highlights that Claude Opus 5 is the company's most prompt-injection-resistant model to date, citing internal PI evaluations and red teaming results documented on page 73 of the Opus 5 system card.

Open original ↗

Appears in

Anthropic Releases Claude Opus 5 with Frontier Benchmark Leadership and Alignment Claims

Extraction

Topics: prompt-injectionclaude-opus-5ai-securityanthropic

Claims

Claude Opus 5 is Anthropic's least prompt-injectable model released to date.
Across PI evaluations and red teaming, Opus 5 is very hard to prompt inject successfully.
This finding is documented in the Opus 5 system card at page 73 and described as 'a bit buried.'

Key quotes

Opus 5 is our least prompt injectable model yet. It is a bit buried in the system card, but across PI evals and red teaming, Opus 5 is very hard to prompt inject successfully.