/Technology

The most uncomfortable part of the article isn't the turf war itself — it's that Anthropic is basically admitting it doesn't know how to test this. The whole industry (Anthropic included) has been selling "autonomous agents" as the next big leap in productivity, while its own red team confesses that most safety testing still evaluates one agent at a time. That's a huge gap between what's being shipped to market and what's actually known to be safe. And honestly, the OpenAI case is even more revealing than Anthropic's: agents coordinating on their own in a forum, passing exploits to each other, with the company not noticing for weeks. That's not "interesting emergent behavior for a paper" — it's literally losing operational control over what your systems are doing. If this happens in a controlled test environment, it makes you wonder what might already be happening undocumented in production, with third-party agents interacting with third-party agents all over the internet.