This week, Anthropic published its September 2026 threat intelligence report featuring a detailed account of how its AI has been misused across surveillance operations, influence campaigns, and data theft. One finding stands out for anyone working with sensitive environmental data: third-party AI labs were caught routing user conversations through Claude and using those exchanges as training data. Those sessions contained names, email addresses, and corporate data from hundreds of end users. The users had no idea.
The report is worth reading given the important question it raises is about the AI tools your organization is already using.
Question the architecture
When an environmental consultant uploads a Phase II report to a general-purpose AI tool, where does that document go? Does it persist in the system? Can it be used to train future models? Could it surface in another user's session?
Most AI providers publish privacy policies. Fewer are transparent about how data flows through their infrastructure, and fewer still have built their systems from the ground up to prevent cross-contamination between clients.
The Anthropic report documents cases where sensitive corporate data was relayed through third-party routing services without users' knowledge. The data moved through systems that were never designed to keep it isolated.
Environmental site data carries the same risk. A Phase I ESA for a pending acquisition. A remediation cost estimate tied to active litigation. A regulatory submission that has not yet been filed. These are documents you want traveling only through infrastructure you control.
Why this matters for environmental work
Environmental site data is among the most legally sensitive information an organization holds. It touches regulatory liability, transaction value, litigation exposure, and insurance coverage, often simultaneously.
Contaminant history that surfaces can shift liability allocation in a transaction. A remediation cost estimate shared with the wrong party can affect settlement negotiations. A regulatory submission reviewed by an AI system that logs conversations may create a record you did not intend to create.
The Anthropic report describes misuse at scale: surveillance platforms, influence operations, data theft. While environmental professionals are not the primary targets of those operations, the underlying vulnerability is the same. Providing data to a system and infrastructure you do not fully understand.
The Statvis approach
Statvis was built for environmental professionals who work with confidential site portfolios. Every document uploaded to a site stays within that site's isolated environment. Our system does not train on client data. Every response cites its source to allow you to verify what the AI read, and confirm that it read only what it should have.
When a lawyer asks how a finding was reached, the answer is a citation to a specific page in a specific document, not a confidence score from a model trained on data you cannot audit.
Confidentiality is built into the architecture of Statvis.
What AI confidentiality requires
Environmental professionals understand chain of custody for physical samples. The same principle applies to documents.
Confidentiality in an AI system means more than a checkbox in a settings menu. It requires:
- Isolated data handling. Each client's documents stay within their own environment. No cross-client retrieval, no shared indexes.
- No training on client data. Documents uploaded to the platform are never used to improve the underlying model for any client.
- Verifiable retrieval. When the system surfaces a finding, it cites the exact source document and page. You can verify what the AI read and confirm it did not fabricate a reference.
- Access controls that match your governance requirements. Who can see which sites, which documents, which findings defined by your organization, not by the platform's defaults.
These are architectural decisions made before the first document is uploaded to the system.
Anthropic's full threat intelligence report is available at anthropic.com.