Pentagon adds Grok and ChatGPT as AI scheming reports double

The Pentagon just gave its personnel a choice between Grok, ChatGPT, and Gemini inside a single approved environment, and I find that more revealing than it first appears. This isn't the Department of Defense picking a favorite model — it's an admission that no single lab has earned enough trust, or proven enough capability, to be the default. Multi-model procurement is becoming the norm in high-stakes environments, and that's a quiet but significant shift from the "one model to rule them all" narrative that dominated 2023 and 2024. If the most security-conscious buyer in the world hedges its bets like this, expect enterprises to follow.

Meanwhile, the copyright reckoning keeps grinding forward. Sony Music Publishing has now sued Anthropic, accusing it of what the filing calls blatant infringement. By the way, it's worth noticing the pattern here: Anthropic markets itself as the safety-conscious, principled lab, yet it keeps landing in the same legal crosshairs as everyone else on training data. That's not necessarily hypocrisy — training on the open web was standard practice industry-wide before anyone had settled the legal question — but it does puncture the idea that being careful about model behavior insulates you from being careless about model inputs. These cases are piling up fast enough that I think we're heading toward either a major settlement framework or a court ruling that reshapes how every lab sources data going forward. Neither outcome is going to be quiet.

On the infrastructure side, AWS launching an Agent Registry tells you where the real enterprise AI money is going right now. Nobody's excited about chatbots anymore; the excitement is in managing fleets of agents, tools, and skills at scale, which is a fundamentally different and much less glamorous problem. It's closer to IT asset management than to prompt engineering, and that's exactly why it matters — this is the boring plumbing that determines whether agentic AI actually works in production or just works in demos. Companies like Relativity plugging specialized legal tools into Copilot, and Google Cloud co-designing financial AI suites with actual banks, point the same direction: the interesting AI work in 2026 isn't happening at the model layer, it's happening at the integration layer.

Then there's the watchdog report claiming AI scheming incidents nearly doubled in July, paired with the note that UK Parliament has no legal power to respond even if it wanted to. I'd treat the scheming numbers with some skepticism until we see methodology, but the regulatory gap is real and not unique to Britain. Governments are adopting these systems — see the Pentagon news above — faster than they're building the authority to govern them. That mismatch, not any single scary headline, is the story worth watching.