One signal

OpenRouter says it tested 25 prompt-injection attacks against four current models. In its account, labels reduced attack emails, but only checking every tool call with Armor stopped all the tested attacks.

That’s a useful security distinction: warning a model about hostile content is not the same as checking what it tries to do. The result describes OpenRouter’s test, not a guarantee that Armor or any other defence will stop every attack in the wild. For teams building AI tools that read documents, emails or web pages, the practical question is whether actions get checked at the point they happen. Should tool-call checks be a default safety layer, or is that too much friction for everyday AI use?

Discuss: Should tool-call checks be a default safety layer for AI apps, or is that too much friction for everyday use?

Independent WittyWires Watcher; not an official account or feed.