AI Pulse · Frontier Labs Watch

Anthropic is reportedly working on a way for enterprise customers to keep required AI monitoring data inside their own infrastructure, rather than having it retained by Anthropic.

This follows a policy change in June that made 30-day retention for security monitoring mandatory on Anthropic’s most capable models. Anthropic’s reason was cyberattack detection. For many companies, though, the issue is not whether monitoring is sensible. It is where that data sits, who can access it, and whether it conflicts with internal security rules, customer contracts, or regulated-data policies.

The reported system is being developed with input from more than 100 customers, including Salesforce. There is no release date, and it does not solve the immediate problem for companies that need zero retention. If you are currently negotiating a Claude enterprise agreement and your policy requires no data retention at all, that requirement still appears to be in force.

The practical lesson is simple. Don’t treat an AI vendor’s data policy as a footnote in procurement. Ask what is retained, for how long, where it is stored, whether your team can control it, and what changes when you use the vendor’s most capable models. Those answers can differ materially by model tier.

This is the kind of thing we build into an AI command centre, so leaders can see which tools are approved, what data rules apply, and where policy changes create operational risk.