Key points
- Three large companies are limiting Anthropic's models
- Palantir wants a zero data retention guarantee
- Nvidia routes proprietary work to its own models
- Anthropic already changed the policy behind the reported concern
Palantir (PLTR), Nvidia (NVDA) and Booz Allen Hamilton (BAH) are each drawing a line around Anthropic's models when proprietary data is involved. The concern is not model quality. It is what Anthropic might retain after a customer sends data or intellectual property into the system, The Information reported on Monday, citing people familiar with the matter.
The lines are different. Palantir has asked Anthropic for a zero-data-retention guarantee before it will offer the models inside its software. Nvidia uses Anthropic for less sensitive internal work and sends proprietary tasks to Nemotron, its own model family. Booz Allen has barred staff from using Anthropic's commercial model on cybersecurity projects involving proprietary data. None of the three companies confirmed the reported arrangements.
The policy behind the reported concern
The dispute traces to a change Anthropic made when it rolled out Fable 5. The company dropped zero data retention for that model and moved to a 30-day window for usage logs, on the reasoning that the worst misuse is only visible across sessions. Its argument was that scanning each interaction on its own and deleting it immediately does not catch the pattern. Anthropic says it has never trained on enterprise data without explicit permission.
That explanation did not solve the problem for regulated customers. Banks, hospital systems and defense contractors operate under rules about where their data sits and who can read it, and a 30-day log held by a vendor is a compliance question before it is a security question.
Anthropic's response, and what it does not settle
Anthropic announced a response on September 1, two weeks before the restrictions were reported. Enterprise Frontier Safeguards lets a customer keep the activity data used for misuse detection in its own Amazon S3, Azure Blob Storage or Google Cloud Storage account, "under their own encryption keys, access policies, and audit logging." Anthropic describes it as combining "the privacy of zero data retention (ZDR) with state-of-the-art safeguards for detecting misuse."
The review step is the part aimed squarely at Palantir's objection. Anthropic says the monitoring is automated, with "no Anthropic human review required," and that when something is flagged the signal goes "directly to the customer and their people take it from there." The company says it developed the system with more than 100 customers, naming Goldman Sachs, Morgan Stanley, Citi, Bank of America, Wells Fargo, Comcast, KPMG, Mastercard, Salesforce and Visa. It does not charge for it.
Enterprise Frontier Safeguards is rolling out in phases, and Anthropic says it is aiming for broad availability later this fall. The restrictions reported on September 14 fall in the window between that announcement and its delivery.
Booz Allen and Palantir both sell AI systems into government and regulated industries, and both resell other companies' models inside their own platforms. Nvidia sits on both sides: Nemotron competes with the models it is restricting, and chief executive Jensen Huang called cybersecurity AI's next major market last week.
Palantir closed at $173.31 on Monday, up 3.6%. Booz Allen closed at $79.11, up 4.2%. Nvidia closed at $210.96, down 3.4%.
