Iran Used Claude to Target US Navy Ships. Here's the Jailbreak Pattern Nobody Caught

Automated excerpt

That's an architecture problem, not a "Claude needs better training" problem. The model never sees the whole plan in one prompt. It doesn't require the model to reason about anything, the pattern is caught structurally. One honest caveat: Sentinel scores each request independently, same as the model itself does. Sources Anthropic Says Iran Used Its American AI Model to Target U.S.

Selected automatically from source text; not independently written or fact-checked. Read the original for full context.