@scalzi@mastodon.social The fun bit here is that this is not Anthropic‘s fault for supplying an LLM that sometimes messes up things. We all do. It‘s Anthropic‘s fault for supplying an MCP protocol that is one big security leak. When it was new I posted here that a „tool usage for AI“ protocol that doesn‘t force the tool provider to provide, for every API, whether it is safe (only reads) or dangerous (changes stuff) is madness. But the protocol is simple and easy to implement, so the world is going for it.