Coverage
Anthropic / Claude
Govern Claude without slowing anyone down
Claude Enterprise puts a capable assistant next to your most sensitive work, and Claude Code puts one inside your repositories. SAF3AI reads both through Anthropic's own admin and compliance interfaces — no proxy, no browser extension, no change to how your teams work.
What you see
The telemetry we pull
Compliance Activity Feed
Every organisation-level action in Claude — conversations started, files uploaded, projects created, members added, settings changed — as a continuous read-only feed.
Conversation content
On Claude Enterprise, the full chat, file and project content behind each activity event, so a data-loss finding points at the actual prompt rather than a timestamp.
Usage and cost reports
Token usage and spend broken down by workspace, API key and model, from the Claude Console Admin API — the input FinOps needs to attribute cost to a team.
Organisation inventory
Users, workspaces and API keys, with who owns which key. Every long-lived key is a non-human identity, and this is the register of them.
Claude Code telemetry
Developer coding-assistant activity, so AI-written code is attributable back to a person, a repository and a session.
Model and tool usage
Which models and tools each workspace calls, how often and with what volume — the baseline anomaly detection scores against.
How it connects
From zero to first signal
- 1
Choose the credential that matches your Anthropic tenancy
A Claude Console organisation uses an Admin API key. A Claude Enterprise organisation additionally provisions a Compliance Access Key, which is the only credential that reaches conversation content.
- 2
Provision it read-only
The Compliance Access Key only ever needs read scopes — read:compliance_user_data and read:compliance_org_data. SAF3AI never requests, and cannot use, the delete scope.
- 3
Paste the key into the connector
Add the credential in Integrations. Keys are encrypted at rest, scoped to your tenant and redacted everywhere they surface in the product.
- 4
Backfill, then stream
The connector pulls recent history so you start with a populated picture rather than an empty dashboard, then polls continuously with a durable watermark so nothing is missed or double-counted across restarts.
- 5
Signals join the graph
Claude workspaces, users, API keys and models become entities in the Context Graph alongside every other surface — so a Claude finding can be correlated with a Copilot one about the same person.
Content collection requires a Claude Enterprise organisation and a Compliance Access Key. A Console-only tenant still gets the full activity feed, usage, cost and inventory — just not conversation content.
What it catches
Risks specific to this surface
Sensitive data pasted into chats
Source code, customer records, credentials and regulated data going into prompts. Validated detection — Luhn-checked card numbers and provider-signature secrets, not a regex that flags every 16-digit string.
Orphaned and over-scoped API keys
Keys that outlived the project or the employee who created them. The inventory ties every key to an owner and a workspace, so the ones with nobody behind them stand out.
Unsanctioned personal accounts
People using consumer Claude alongside the corporate tenant. Shadow AI discovery catches what the Enterprise connector cannot see, and coaches them back to the governed workspace.
Project and file oversharing
Knowledge bases and project files reachable by more of the organisation than intended — the Claude equivalent of the overshared SharePoint site.
Prompt injection through uploaded files
Instructions hidden in an uploaded document that redirect the assistant. Includes Trojan-Source and invisible-character detection, not just literal text matching.
AI-written code entering the codebase
Claude Code output landing in repositories without review. Pairs with Code Security to scan what was generated for vulnerabilities and secrets.
See Anthropic / Claude in your own tenant
Connect this surface in a pilot and get a mapped inventory, a scored risk list and the attack paths that actually reach your data.