Anthropic will let banks keep Claude's misuse logs in their own cloud
Enterprise Frontier Safeguards stores monitoring data in infrastructure the customer controls, with Anthropic running risk evaluation without seeing the telemetry. It resolves the objection that has blocked regulated buyers from agentic deployments.
Anthropic has introduced Enterprise Frontier Safeguards, an architecture that lets a customer keep the safety monitoring logs Claude generates inside cloud infrastructure the customer controls, rather than inside Anthropic's. The company still performs risk evaluation; it does so without holding the underlying telemetry.
The feature shipped alongside Claude Fable 5.1 on 1 September. It is aimed squarely at banks, and it addresses the objection that has done more to slow regulated adoption of agentic AI than capability ever did.
The objection it answers
Misuse monitoring requires reading what users send. For a frontier lab, that is a safety necessity. For a bank, it means the contents of employee prompts — draft credit memoranda, positions, client names, deal codenames, internal investigation material — sitting in a vendor's logs.
That is a problem no amount of contractual assurance fully solves, because the risk is not only that the vendor misbehaves. It is that the vendor is subpoenaed, breached, or acquired. A bank's supervisor asks where material non-public information is stored, and "a third party retains it for safety purposes" is a difficult answer in most jurisdictions.
The usual workarounds were worse. Buyers either disabled the deployments that would carry sensitive material, restricting AI to low-value work, or negotiated bespoke arrangements that took months and produced weaker monitoring.
How it is meant to work
Under EFS, the monitoring data is written to storage the customer owns and controls entirely. Anthropic evaluates risk against that data without taking custody of it. The customer keeps its logs, its retention schedule and its own legal position over the material, and the safety function still runs.
This is the same shape as the confidential-computing arrangements that made cloud acceptable to regulated industries a decade ago: not a promise about vendor conduct, but an architecture in which the promise is not required.
What Anthropic has not published
The details that determine whether this satisfies a regulator are not yet public. Anthropic has not described what signal it receives when evaluation runs against customer-held data, and any such system returns something — the question is how much can be inferred from it.
Nor has it said what happens on a positive finding: whether Anthropic can see the flagged material, whether it can suspend access unilaterally, and what it is obliged to report to authorities when a customer's own logs indicate serious misuse. A safety architecture in which the vendor cannot see the evidence is also one in which the vendor cannot easily act on it.
There is no independent attestation, no published third-party audit, and no named launch customer.
Why the competitive timing matters
Anthropic holds roughly 43.5 percent of US enterprise AI spending, against OpenAI's 39.7 percent and Google's 6 percent, according to Ramp data from August cited by TechCrunch — figures Google notes exclude large strategic deals.
That lead was built in exactly this constituency, and it is being contested. Google and Accenture announced a joint deployment unit this week; Microsoft, OpenAI and Amazon all now run forward-deployed engineering practices. Deployment capacity is becoming the axis of competition, and data residency is the specific blocker EFS removes for the most valuable buyers in the market.
Runs the newsroom. Rename this profile in the studio to your own byline.
Related
Every weekday, the AI stories that moved money or shipped code.
No cross-posting, unsubscribe anytime. See all newsletters