Personal data stops before the model sees it.
Arbor Guard sits between your people, your applications and the AI models they use. It finds personal data and secrets in a prompt and, by the rule you set, lets it through, warns, redacts it, swaps in a token or blocks the request. Prompt text is held in memory for one request and is never stored.
- Applications and staff chat
- Redact, tokenise or block
- Prompt text never stored
Runtime protection, not governance: the register, the assessments and the evidence are Arbor Govern. Built to run hosted in a region you choose, or inside your own Kubernetes cluster.
✓ luhn
✓ verhoeff
✓ 0 values kept
Someone pastes a customer's card number into an AI chat. The model receives the sentence without the number, and you receive a record that it happened, without the number either.
Wherever a prompt leaves, one policy applies
Your applications call seven model providers through one gateway. Your people use AI chat in a browser. Your network already has a proxy. Each is a way in, and the same rules decide what happens at all of them.
A gateway in front of the model API
Change one line, the base URL, in the OpenAI or Anthropic SDK your application already uses. Streaming answers, attached files and images are checked, with text read out of images and PDFs by OCR. Coding tools that take your own key, such as Claude Code, Continue, Cline and Aider, go through it too.
- OpenAI
- Anthropic
- Google Gemini
- Azure OpenAI
- Amazon Bedrock
- Vertex AI
- Sarvam
A managed browser extension
Checks a prompt on paste, on submit, on a file picked or a file dropped, before it is sent. Word, Excel, PowerPoint and PDF files are read inside the browser. Rolled out to Chrome, Edge and Firefox through Google Admin, Intune or Jamf.
- ChatGPT
- Claude
- Gemini
- Copilot
- Microsoft 365 Copilot
- Perplexity
- DeepSeek
A check your web proxy calls
An ICAP server your existing proxy sends AI traffic to, covering the ChatGPT, Claude, Perplexity and DeepSeek web apps and the OpenAI, Anthropic and Gemini APIs. Your proxy has to decrypt that traffic first. Tested with Squid.
Coverage from the logs you already keep
Reads Zscaler, Netskope-style W3C and Palo Alto URL-filtering logs inside your network and shows which AI services people reach that nothing is inspecting: desktop apps, phones, coding tools. Only a count per AI service leaves, never a path, a user or an address.
Five answers, and the strongest one wins
When two rules disagree about one prompt, block beats tokenise, tokenise beats redact and redact beats warn. A rule never weakens another by being written later.
- 01
Allow
The request goes through and an event is recorded.
- 02
Warn
The request goes through, the person or the application is told why, and an event is recorded.
- 03
Redact
The value is replaced before the prompt leaves, and the model never sees it.
- 04
Tokenise
The model sees a token in place of the value. At the gateway the real value is put back into the answer, including inside a streamed reply.
- 05
Block
The request is refused with the reason, and nothing reaches the model.
What it looks for
Patterns with their published check digits, the same detectors Arbor Lens uses, plus a name list. An optional model finds names and addresses a pattern cannot.
- Indian identifiers
- Aadhaar with its Verhoeff check, PAN, GSTIN, IFSC, mobile numbers, PIN codes, voter ID, driving licence, UPI ID
- Payment and banking
- Card numbers with their Luhn check, IBAN with its mod-97 check
- Contact and people
- Email, IP address, names, dates of birth, postal addresses, US Social Security numbers
- Secrets
- AWS access keys, GitHub tokens, OpenAI, Anthropic and Google API keys, Slack and Stripe tokens, private keys, JWTs
- Prompt attacks
- Prompt injection and jailbreak patterns
- Your own terms
- Project names, client names, anything you list, matched as literal phrases
Rules you can roll out without breaking anyone
A rule per AI system and per group
Each detector gets an action for prompts and another for answers, per AI system and per browser group. New systems can start in warn-only mode for a set number of days before they enforce.
Signed policy, reviewed before it ships
A policy change is staged and reviewed before it is published. Gateways receive it as a signed bundle and refuse one that is older, expired or unsigned, keeping the last good bundle instead.
Break-glass that takes two people
One admin asks for a time-boxed exception with a reason, a different admin approves it, and every use is recorded and announced by email, Slack or Teams.
Open or closed, your choice
Detection runs within a time budget. If it runs out, each AI system either refuses the request or forwards it and records that it did, whichever you chose for that system.
A record of what happened, never of what was said
Prompt and answer text live in memory for one request. A prompt still goes to the model provider you chose, and a tokenised value goes to the vault that issues its token. What Guard itself writes down is this, and only this.
Which detector fired, which action was taken, how many values, and a keyed hash of the user. Never the prompt, never the answer, never the value found.
Events go onto a per-tenant SHA-256 hash chain, so a changed or deleted row is found when the chain is verified.
A signed evidence file is sent daily to Arbor Govern, where it sits beside the AI system it belongs to.
Retention is set per tenant within the limits of its region, and every purge is itself written to the chain.
Stated plainly, so nobody demos something that does not exist
A security buyer who finds an unstated gap stops believing the stated claims too. These are the ones you would otherwise find in week three of an evaluation.
Agent tool calls and MCP are not inspected
Agents are covered when they call a model through the gateway. Traffic between an agent and its tools, including MCP servers, is not checked today.
Some AI lives where the extension cannot reach
Gemini inside Gmail and Docs, Copilot in the Edge sidebar and Copilot inside the Office desktop apps are not covered. Neither are the completions Cursor and GitHub Copilot send to their own vendors.
The extension installs by managed rollout
It is not yet listed in the Chrome, Edge or Firefox stores, so it reaches browsers through Google Admin, Intune or Jamf. The Firefox build passes its checks but has not yet run in a live Firefox.
The network proxy checks requests only
Answers are not checked on the proxy path, and a proxy cannot put a token back into an answer, so tokenising there is a replacement. Traffic the proxy does not decrypt is never seen.
Five provider routes await a live call
The Gemini, Sarvam, Azure OpenAI, Bedrock and Vertex routes follow each provider's published request format and have not yet carried a call from a customer deployment. Hosted regions have not yet carried customer traffic, and tokenising through a live Arbor Vault is untried; without a vault, Guard issues its own tokens for the one request.
Detection reads identifiers and English
Names and addresses are found in English text. Guard makes no claim about names or addresses written in Indian scripts, and there is no SIEM export yet.
By traffic and seats, never by surprise
Licensed on request volume and browser seats, and quoted on request. Billing never blocks traffic: only a policy blocks a call. Self-hosting inside your own cluster is available on the largest licence.
Does Arbor Guard store prompts?
Arbor Guard does not store prompt or answer text. A prompt is held in memory for the one request it belongs to, and the record kept afterwards names the detector, the action and a count, never the value that was found.
Which AI providers does it work with?
The Arbor Guard gateway speaks to OpenAI, Anthropic, Google Gemini, Azure OpenAI, Amazon Bedrock, Vertex AI and Sarvam. An application changes one line, the base URL, in the SDK it already uses.
What happens to an answer when a value was tokenised?
Tokenised values are put back into the answer at the gateway before your application sees it, including inside a streamed answer where a token arrives split across pieces. The model only ever saw the token.
Do staff have to change how they use AI chat?
Staff keep using the chat sites they already use. The extension checks a prompt when it is pasted or submitted and a file when it is attached, and either lets it through, warns, redacts or stops it, by the policy for that person's group.
Is Arbor Guard part of Arbor Govern?
Arbor Guard is a separate product on its own licence. Its policies, events and approvals are managed from the same console as Arbor Govern, so a team running both signs in once and sees runtime events beside the AI systems they belong to.
What happens if detection cannot finish in time?
Each AI system is set to fail closed or fail open. Closed refuses the request when detection cannot finish in time; open forwards it and records that it went unchecked.
Last reviewed . General information, not legal advice. Arbor Guard is a separate product from Arbor Govern and tokenises through Arbor Vault where one is deployed, and every product is compared here.
Start in warn-only. Watch for a week.
Point one application or one browser group at Guard with every rule set to warn. A week of events shows what your people and your systems actually send, before a single request is changed.