The most important AI story of the week was not a model release. It was the argument over who gets to decide what a powerful model is allowed to do. Anthropic's clash with the Pentagon, OpenAI's rapid agreement with the same institution, and Perplexity's new agentic Computer all pointed at the same deeper shift: AI policy is no longer a back-office document. It is product infrastructure.
The Anthropic dispute made that painfully visible. TechCrunch reported that Anthropic resisted giving the Pentagon unrestricted access to its AI systems, especially for mass surveillance and autonomous weapons, while the government threatened extraordinary pressure in response. OpenAI then announced its own Pentagon deal with technical safeguards, and followed with more details after the obvious public question: if two companies claim similar red lines, why could one sign and the other could not?
My read is that the industry is entering the contract era of AI. The real capabilities of a model will increasingly be shaped not only by weights and benchmarks, but by deployment venue, monitoring stack, approval workflow, audit trail, and the power relationship between vendor and customer. Those are product choices, but they also become political choices when the customer is a state actor.
Perplexity's Computer announcement from Ars Technica belongs in the same frame. A tool that assigns work to other AI agents sounds like convenience, but it also creates a new chain of accountability. If one agent plans, another searches, another writes, and another executes, the user needs more than a chat transcript. They need a map of responsibility.
That is why I think March opened with a useful warning. The next AI platforms will not be judged only by what they can do. They will be judged by whether their boundaries survive contact with powerful customers, impatient users, and complex workflows. Guardrails that exist only as marketing language will not be enough. The guardrails have to be part of the machine.
References
- Anthropic accuses Chinese AI labs of mining Claude as US debates AI chip exports, TechCrunch, February 23, 2026.
- Anthropic won't budge as Pentagon escalates AI dispute, TechCrunch, February 24, 2026.
- Perplexity announces "Computer," an AI agent that assigns work to other AI agents, Ars Technica, February 26, 2026.
- OpenAI's Sam Altman announces Pentagon deal with 'technical safeguards', TechCrunch, February 28, 2026.
- The trap Anthropic built for itself, TechCrunch, February 28, 2026.
- OpenAI reveals more details about its agreement with the Pentagon, TechCrunch, March 1, 2026.