■ The Big Story

Anthropic wants to prove which AI answered your question

Anthropic introduced FlashProve in research announced on X, saying it lets developers check whether a response came from a particular AI model. The system builds on established cryptographic techniques. Think of it as a tamper-resistant receipt for a model’s identity, not a guarantee that its answer is right. That distinction matters when a supplier handles work through several AI services.

TechCrunch reports that Anthropic is working on FlashProve as an open-source system. It is designed to provide a mathematical guarantee linking an output to a specific model. For a buyer, that would make a vendor’s promise about which model handled sensitive work easier to check. Answer quality, data handling and responsibility for errors would still need their own contract terms.

Sources: [Anthropic on X] [TechCrunch]

■ Frontier Watch

■ The Street Says

The argument is less about whether AI can act and more about who controls it, what vendors can be trusted to handle, and whether changing models disrupt paid work.

■ Repo of the Day

■ Two Steps Ahead

Over the next 30–90 days, Anthropic’s FlashProve gives procurement teams a precise question for AI suppliers: will a contract let the customer verify which model answered? Cohere’s Model Vault puts a separate question beside it: what happens to prompts and outputs? Keep those promises distinct when comparing vendors; proof of model identity cannot substitute for data-handling terms.

CrowdStrike’s emergency stop makes shutdown rights a practical requirement for any AI worker given network access. Teams using Google Antigravity for drug research face a different handoff: who approves a proposed molecule before lab spending begins? JustVugg / colibri is relevant to technical teams calculating whether existing equipment changes the cost of running models themselves, not a reason for other businesses to rewrite software contracts.