Skip to content

Policies and personal data

You decide which personal data is masked before it reaches a model, and which tools may run without asking. Both apply to every chat and every agent. This page covers the personal data policy, the tool rules, and what people see when a rule stops something.

Before you start

  • For personal data: the PII / Privacy module is on (Settings → Modules; see Modules). Its pages are under Privacy (PII). Without it, nothing is masked, with one exception: see When the module is off.
  • For tool rules: Agents → Governance. It needs no module.

Personal data

The personal data filter finds names, e-mail addresses, phone numbers, account and card numbers, places and similar data in what people send, and masks it before the request reaches a model. It runs only as your policy says: you choose which data, for which models, and what happens when the filter cannot run.

Detection is not always right

How well the filter works depends on the installed profile, the language and the structure of the content. Tables, lists, code and short fragments need particular care. It can miss personal data, and it can mask words that are not personal data. Try it on your organisation's typical content before you rely on it, and tell people not to paste what they may not share.

The policy Fadenstack starts with

With the module on, the default policy:

  • protects requests to cloud providers; requests to your own machines' models are not scanned;
  • redacts: personal data is replaced, and the model never sees it;
  • blocks a request to a cloud provider when the filter cannot run;
  • reads English text and names no types of its own: it looks for the standard types the installed detector finds, with that detector's thresholds. These are PERSON, EMAIL_ADDRESS, PHONE_NUMBER, CREDIT_CARD, IBAN_CODE, IP_ADDRESS, US_SSN, LOCATION, MEDICAL_LICENSE and URL, and with the older presidio-legacy profile also NRP (nationality, religion, political group), which the local detectors do not find. The policy form lists them under Entities when none are picked.

Before 0.6.0 the default named eleven types itself, NRP among them, with a threshold for the older profile. The local detectors refuse both, so with one installed every request to a cloud provider was refused. An install that still has that default as it was set up gets the new one when it starts; a default you changed stays as you saved it (see Adopt an installed detector profile).

When the module is off

A policy needs the module to run. What happens without it depends on where the policy comes from:

  • A policy saved for your organisation (Settings → Modules → PII / Privacy, tenant policy): a missing module counts as a filter that cannot run, and If detection fails decides. With block, the default, chats to the models the policy covers fail with "A PII policy is set for this workspace, but the PII module is not running" and nothing is sent. Start the module, or switch the policy off if you do not need it. With allow the request goes out unscanned, and the log says so.
  • The default policy only (nobody saved a policy): not applied. Every install has it, including those that never switched the module on, and their chats keep working unmasked.

Switching the module off therefore does not silently remove a policy you saved. Be aware that if you rely on the default policy and later switch the module off, requests to cloud providers are no longer masked: save a policy for your organisation if that must not happen.

While that is the case, administrators see it at the top of every admin page, as an error that cannot be hidden: "PII is off: requests to cloud models go out with personal data unmasked." It appears once the module, on before, is off, a cloud model is enabled and no policy is saved for your organisation; and it goes when the module is on again (faden module enable pii), when no cloud model is enabled any more, or when the default policy no longer scans cloud models (then sending them unmasked is what you chose).

More about dates

Dates (DATE_TIME) are not on the list. Add them if you need dates of birth covered; plain numbers in a message may then be masked as dates too.

The older presidio-legacy profile supports English. The local multilingual profiles support English (en), German (de), French (fr) and Spanish (es). The form lists the languages and categories available from the running service. A successful check does not confirm protection for an unlisted language.

Change the policy

The default policy applies to the whole installation:

  1. Open Settings → Modules, find PII / Privacy, and open its Module settings.
  2. Change the Default PII Policy (fields below).
  3. Select Save.

Warning

Saving the default policy restarts the gateway to apply it. Replies being written at that moment stop and can be tried again. Save outside busy hours.

A tenant policy replaces the default for one tenant (your organisation is the tenant default unless your installation uses others):

Privacy (PII) → Tenant Policies: the Tenant ID box and Load Tenant

  1. Open Privacy (PII) → Tenant Policies, enter the Tenant ID and select Load Tenant.

    A tenant policy: Detection (Presidio) with the language, the score threshold and the entities, Clear Override and Save, the switch Let chats and agents change the PII mode, and the Effective Policy Preview

  2. Change the fields under Tenant Override. Effective Policy Preview shows what requests will get.

  3. Select Save. Clear Override goes back to the default policy.

Adopt an installed detector profile

Installed detector shows the model, content locale pack and execution mode. Only the pack’s installed languages are offered. A locale change requires review of each affected saved policy.

After your operator changes the detector or its entity filter set, review the default policy and each tenant override that should use it:

  1. Check that the installed profile is available. An unavailable-service notice does not confirm active protection.
  2. Select Use installed profile defaults under Detection (Presidio). This selects the installed profile and Use profile thresholds, and removes the categories the installed detector does not find; it keeps your other categories and outgoing-request choices. The form warns when a policy's threshold or a category does not fit the installed detector: until it is adopted, requests the policy covers are refused.
  3. Choose Language. Remove categories shown as unavailable for that language, then add the categories you need under Entities. Adopting a profile does not automatically select every available category.
  4. Choose whether to protect cloud requests, local-model requests or both. Keep If detection fails at block when a protected request must not be sent without a successful check. Save the policy. On a server without a GPU, a request with a long text can take longer than the check allows: see Local PII on CPU or GPU.

Changing the language can make a previously selected category unavailable. Remove or replace it explicitly; the service refuses unsupported coverage instead of silently ignoring that category.

Extended categories and their limits

When your operator enables the extended filter set, the available categories include dates, age, gender, postcodes, identity documents, tax and social-security numbers and declared sensitive fields. Select the categories your policy needs; availability alone does not enable them.

Health, ethnic origin, political opinions, religion or philosophical beliefs, union membership, genetic and biometric data, sex life, sexual orientation and criminal records are recognized in supported explicit fields. This does not provide general detection of every sensitive statement in ordinary conversation, nor a guarantee of GDPR compliance. Test your organisation's content and field formats.

Numeric or container values under a protected sensitive tool-argument key may be refused because replacing them would change the tool's data type. Non-sensitive values keep their types. Images and audio are not text checks; unsupported protected content follows If detection fails.

More about the fields (the same in both places)

The fields are the same in both places:

Field What it does
Outgoing requests: Protect requests to cloud providers, Protect requests to local models Which requests are scanned.
Mode redact: personal data is replaced with a placeholder such as <PERSON>, and the answer keeps it. tokenize_reversible: each value is replaced with a token such as [PERSON_1], and put back in the answer the person sees. The model sees neither.
If detection fails block: the request is refused. allow: it is sent unscanned. fallback_to_local: it is sent unscanned to one of your own models under the same name, if there is one, and refused otherwise.
Detection (Presidio): Language, Score threshold, Entities The installed profile, language and categories to use. Use profile thresholds uses the installed profile's defaults. A manual higher threshold masks less. Entities left empty: the standard types the installed detector finds. Choose only an advertised language and supported categories; otherwise the check fails and If detection fails decides.
Telemetry: Remove personal data from telemetry, Mode For request traces, which are off by default: whether a trace keeps the request with personal data removed (sanitized) or figures only (metrics_only). Without Enterprise, traces keep figures only anyway.

Chats and agents

The Privacy tab of the chat's side panel: protection of requests to cloud and to local models, the mode and the fail action, telemetry, and the detection threshold and entities

  • A person can make the policy stricter for one chat in the chat's Privacy panel: more types, a lower threshold (it then masks more), local models scanned too.
  • Switching between redact and tokenize_reversible is not a stricter setting but a different one, so a chat may do it only when Let chats and agents change the PII mode is on (Privacy (PII) → Tenant Policies, off by default). Otherwise the chat shows "Your admin hasn't allowed changing the PII mode for this chat."
  • An agent's Privacy setting (Changing the PII mode on its Definition tab) overrides that switch for the agent's conversations: As the tenant's PII policy says, Allowed or Not allowed.
  • What an MCP tool receives is set per server: see MCP servers.

Agents that act on pages and documents run under tokenize_reversible

The browser agent, the Office add-ins and any agent that fills in forms or writes documents work only with real values in their tool calls. Under redact, a name the person typed reaches the model as <PERSON>, and the model can write that placeholder into a form. Under tokenize_reversible, the gateway puts the real value back into the tool call before the app runs it. Set Mode to tokenize_reversible in the policy that covers them, or set the agent's Changing the PII mode to Allowed so that its app can choose it.

In a tool result, the element refs a page agent lists, such as [e50] or [f1.e3], are never masked, in either mode: they name parts of the page, not people, and the agent needs them to act on it.

What the filter found

Privacy (PII) → Dashboard shows counts only: scans, how many found personal data, the kinds found, and the daily volume. Privacy (PII) → Activity Log lists every scan with its mode, kinds and counts, without the text itself.

The dashboard also shows Avg Processing Time and p95 Processing Time; the activity log shows Duration. These describe the PII service's reported processing time, including waiting for detection. They exclude the model's reply time. Older events can have no duration. A detection-rate percentage is the share of checks that found something, not a measurement of how much personal data was successfully removed.

Tool rules

Agents → Governance on the Tool rules tab, beside Tool classes and Decisions: no rules yet, and Add rule

Agents → Governance → Tool rules decides, by name, what happens when a model wants to run a tool:

  1. Select Add rule.

    New tool rule: Name, Tool pattern, Action set to Deny, Message to the user, Category, Priority 100 and Enabled

  2. Give it a Name (for you) and a Tool pattern: a tool name, or a pattern with *, such as mcp.github.*, knowledge_search or *.

  3. Choose the Action: Deny or Ask first (the person approves each call). A rule also covers the tools an agent's app runs itself, such as the browser and Office agents' tools: the app asks every time, whatever the person allowed before, and an app that cannot ask is not offered the tool.
  4. Write the Message to the user: what people are told when the rule applies.
  5. Set the Priority. It orders rules with the same action: when several of them match, the one with the lowest number gives the message.
  6. Save it. Rules apply at once.

Without rules, tools run as the tool mode says.

When several rules match

Deny wins over Ask first, whatever their priority. To leave some tools out of a Deny rule, make its pattern narrower: deny mcp.github.delete_* rather than mcp.github.*. To let a tool run without asking, give it a tool class (see Tool classes and tool modes).

Allow rules from earlier versions

In earlier versions an Allow placed before a Deny or Ask first rule (a lower priority number) let the tool through. From this release a Deny or Ask first always wins, and Allow changes nothing. Rules made with it are listed as No effect: change them to Deny or Ask first, or delete them.

Tool classes and tool modes

Agents → Governance on the Tool classes tab: no classes yet, so every server tool is unclassified, and Add class

Agents → Governance → Tool classes says what tools do, by pattern: read, write, destructive, or As its MCP server says (trust what the tool's MCP server declares about itself). Reaches outside marks a tool that reaches outside the organisation: below Auto a call to it asks, a read tool's too, and the person can allow it for that chat only, not always, since what it is given leaves the organisation. Auto runs it as its class says. A tool nobody classifies is unclassified. An agent's own tools take their class from its Host tools; an agent whose version lists none has every tool unclassified.

People pick a tool mode per chat, and the class decides what runs:

Mode read write destructive unclassified
Off not offered not offered not offered not offered
Read only runs not offered not offered not offered
Ask (every chat starts here) runs asks asks asks
Auto runs runs asks asks

No mode up to Auto runs a destructive or unclassified tool without asking. To let a tool run in Auto, classify it as read or write. A fifth mode, Bypass, is off unless you allow it. Fadenstack's own tools (the knowledge base search, the time, the tool search) are read tools. So is find_tool, the search the agent SDKs bring when an agent allows Local MCP servers (see Agents): it only finds the tools of the MCP servers on the person's computer and switches them on for the chat. Each tool it finds is unclassified and asks first, every time.

The agent SDKs keep five tool names for their own tools, so no app's tool can take one, and the gateway gives each its class unless your rule or the agent's Host tools say otherwise:

Tool Class Reaches the model where the agent allows
find_tool read always
load_skill, read_skill_resource (the person's own skills) read Local skills
web_search read, reaches outside: asks below Auto Local MCP servers
run_skill_script never: skill scripts stay off

Where the agent does not allow it, the tool is held back, whatever its Host tools list.

Tool mode limits

Enterprise

With Enterprise, Agents → Governance → Tool mode limits sets the most a chat's tool mode may be, for Everyone, A team, An agent or A person: choose who it is For, pick the team, agent or person, choose At most and select Set limit. Change a limit in its row, or remove it. Several limits can apply to one chat (everyone's, the person's teams', the agent's, the person's own): the narrowest wins. People see that their chat is capped, and an agent's applications get its limit too. Changes apply at once and are recorded in the audit log.

Without a valid Enterprise licence the limits cannot be applied, so tools in every chat are read-only until the licence is back; a notice for administrators says so. Lift on that notice gives every chat the mode each person picks again; the limits are kept, and apply again once the licence is valid.

Approval limits

Enterprise

With Enterprise, Agents → Governance → Approval limits sets, for Everyone, A team, An agent or A person, how long a grant may last and what the server keeps of the calls it covers. Choose who it is For, pick the team, agent or person, give a Longest grant (minutes) (15 to 720), an Evidence level, or both, and select Set limit. Several limits can apply to one grant: the shortest longest grant and the most recorded level win. The agent's own longest grant and Longest grant (minutes) in Settings → Approvals still apply; the shortest of all of them counts.

Evidence level What the server keeps of each covered call
Decisions The tool, the time, which side ran it, and which grant covered it
Calls That, and its place and arguments after personal data is masked
Results That, and whether it worked: the error, or the first 500 characters of its result after personal data is masked

Without a limit, grants for a time keep Decisions, and working alongside and Bypass keep Calls. A level you set replaces that, also for Bypass. Places, arguments and results are kept encrypted on the server.

A grant takes its longest time and its level when it is made. A shorter limit takes effect at once: running grants it covers that would last longer end at their start plus the new limit, and each such change is recorded beside the grant. A longer limit, a removed limit or a changed level leaves running grants as they are.

Bypass for. Below the limits, Bypass for names whom Bypass is for, on two sides, each on its own: people (persons and teams) and agents. Once a side names anyone, only those on it may use Bypass: a person must be named or be in a named team, and when agents are named too, the chat's agent must be one of them. Everyone it does not name gets Auto at most. The list only narrows: the installation must still allow Bypass, and the agent's highest level must be Bypass.

  • A side stays restricted when its entries go, removed here or deleted with their team, person or agent. A side with nobody left names nobody, so Bypass is for nobody on it, until you select Bypass for any person (or Bypass for any agent), which lifts the side and takes its entries off the list.
  • A narrower list ends the running Bypass sessions of the chats it no longer names, at once, and says how many: the first entry of a side, an entry removed, a side restricted. A wider list ends nothing.

Every change is recorded in the audit log. Without a valid Enterprise licence the limits and the list cannot be applied, so no grant is made while any exist: each call asks. A notice for administrators says so, and Lift on it lets people make grants within the core limits again.

Alerts by mail. With mail set up, administrators and the addresses under its Alerts get a mail:

  • once a day, as one digest, for every switch-on since the last digest: who, which agent and version, on which device, from when until when, how and when they signed in again, and the reason they gave, if any. It goes out at 07:00 in the installation's time zone (the one set under the usage settings; UTC unless set). A day without switch-ons sends none;
  • when the report of a Bypass session's device differs from what the server recorded: at once, once per session.

Agent actions

Enterprise

Agents → Governance → Agent actions shows what agents did and how each call was allowed, newest first:

State Meaning
Asked The server asked the person and has no answer recorded
Once, For this chat, Always The person allowed it, for this call, for the chat, or always (Approved from an app that does not say which)
Declined, Expired The person said no, or did not answer in time
Blocked A rule refused the call
Grant, Bypass A grant or a Bypass session covered it, without a question
Without asking The server ran it without a question: the chat's tool mode allowed it
Ran The device reported running it; a question it asked first is its own row

Each row shows the time, the person, the agent and the version in effect then, the tool, the device, the turn, the model, and whether the row is the server's record or the device's report. Search by person, agent, device, workspace and time (at most 31 days; the last day by default). A workspace is a document, a folder or a site: the search finds the calls a grant covered there. Calls opens a grant's calls with their places, arguments and results, as far as the grant kept them. Export (CloudEvents) downloads the same rows as CloudEvents in JSON Lines, one event per line, for a SIEM to read in; nothing is sent anywhere by itself.

Only people with the role approvals_auditor can open the actions and the export. Every search, export and refused attempt is recorded in the audit log.

Works council (Germany)

The actions are records per person, so using them needs an agreement with the works council (BetrVG §87(1) no. 6). They are kept for analysing incidents, not for judging performance.

Decisions

Agents → Governance on the Decisions tab: filters for Stage, Outcome and Agent, and three tool decisions, Approval required, Expired and Approved, each with what the user was told

Agents → Governance → Decisions lists what the checks decided: when, at which stage (Prompt, Answer, Tool), the outcome (Blocked, Approval required, Approved, Rejected, Expired), what the person was told, the user, the agent and the request. The refused text itself is never stored: only a keyed fingerprint, so the same text refused twice shows the same value, and nobody can read it back.

Running grants

Agents → Governance → Active grants lists the grants that are running. A grant is what a person gives one agent on one of their devices, for a while: either one tool (For a time), or reading and writing in named places (Working alongside), so that what would otherwise ask runs without asking. It never covers destructive or unclassified tools, and never what an Ask first rule or a tool marked Always ask covers. Each row shows the person, the agent, what it covers, the time left and the number of calls used. Revoke ends one from the agent's next call; a call already running finishes. People can end their own in the agent's Settings → Approvals. A grant also ends at its time, and when the agent gets a new version.

Bypass

Bypass is a fifth tool mode above Auto, for long repetitive work a person trusts: in one conversation, destructive and unclassified tools run without asking, for a chosen time. It is off by default. People switch it on in the agent's app, see Tools in the chat.

Allow it

  1. Start Root Mode, which changing these settings needs.
  2. Open Settings → Approvals and turn on Allow Bypass (approvals.bypass_allowed). Turning it on needs Root Mode; turning it off does not. Each change is in the audit trail with its old and new value and whether Root Mode was on.
  3. In the agent's version, set Approval levels → At most to Bypass (see Tools and approvals). Such a version can be published only while Bypass is allowed.

The Starts in level can never be Bypass, so no conversation starts in it. A tool mode limit below Bypass (Enterprise) removes the choice.

Settings → Approvals also holds:

Setting Default What it does
Longest grant (minutes) (approvals.longest_grant_minutes) 120 (15 to 720) The longest any grant may last: for a time, for a stretch of work, or Bypass. An agent's own longest grant may only be shorter.
Keep Bypass records (days) (approvals.bypass_retention_days) 365 (90 to 1095) Bypass sessions and their calls are deleted this long after they ended.
Result summary (characters) (approvals.result_summary_chars) 500 (0 to 2000) Where a grant records results, how much of each covered call's result is kept after personal data is masked. 0 keeps only whether it worked and an error's text. An error always keeps its first 500 characters.

The second sign-in. To switch Bypass on, a person signs in again in the browser. This cannot be switched off. Active grants (and, for holders of approvals_auditor, a session's calls) show how and when the person signed in again: password, OpenID Connect, SAML or directory, with the provider's name. Check this with your sign-in provider before you allow Bypass:

  • An OpenID Connect provider is asked with prompt=login and max_age=0 and must answer with auth_time.
  • A SAML provider is asked with ForceAuthn, and its assertion's AuthnInstant must be later than the request.
  • A provider that answers from an older session is refused: "The sign-in provider did not confirm a fresh sign-in".
  • A directory (LDAP) asks for the password again.

Apps that cannot do the second sign-in (older SDK versions) are not offered Bypass.

What Bypass does not skip. Hidden and refused tools stay so. Deny rules, Ask first rules and tools the agent's author marked Always ask still apply. Tools that act on machines or the physical world are not recognised on their own yet: mark them Always ask or give them an Ask first rule.

What is recorded. Every call that runs under Bypass is recorded on the server: the tool, the time and its arguments after personal data is masked, reads and writes alike. At the end of each turn the person's device reports the calls it ran, and the server compares; any difference is flagged. Bypass itself is recorded before it takes effect, and without the server's confirmation there is no Bypass.

Watch and stop it. Agents → Governance → Active grants lists running Bypass sessions, with any differences, and Revoke ends one (see Running grants).

Switch it off. Turning Allow Bypass off makes Bypass act as Auto at once (within half a minute on every server worker) and ends running Bypass sessions, recorded as ended by the switch.

Who may read what ran. Only people with the role approvals_auditor. Nobody has it at install, and being an administrator is not enough. Every read and every refused attempt is in the audit trail. The console's list shows the tool, the time, which side ran it (device or server), whether the device reported it, differences, and, as far as the grant kept them, the place, the arguments and the result. They stay encrypted on the server and are decrypted only for the read. With Enterprise, Agent actions searches them across sessions.

Works council (Germany)

Records per person need an agreement with the works council (BetrVG §87(1) no. 6). The records are for analysing incidents, not for judging performance.

What people see when something is stopped

What happened In the chat
A tool rule denied a tool The tool line says Not run. The model is told, with the rule's message, and carries on without it.
A tool needs approval "… wants to run", with Approve and Decline. Without an answer within five minutes, the call is not run (Expired).
The personal data filter could not run, and the policy blocks The reply fails with "PII service unavailable and fail_action=block." Nothing was sent to the model.
The policy does not fit the installed detector (a threshold for the older profile, or a category it does not find), and the policy blocks The reply fails with "The PII policy does not fit the installed PII detector (…), so nothing was sent." Adopt the installed profile in the policy (see above).
A policy you saved applies, but the PII module is not running, and the policy blocks The reply fails with "A PII policy is set for this workspace, but the PII module is not running, so nothing was sent."
A prompt was refused (content policies, Planned) The message is marked Not sent, with the policy's message or "This was stopped by your organisation's policy." Nothing is sent to a model or stored.
An answer was refused (content policies, Planned) The answer stops and is marked Answer withheld.

Applications using the API get the same outcomes as errors they can handle; the details are on the Developers site.

What Core includes, and what is planned

Check Core Enterprise
Who may sign in and use an agent Accounts, roles, agent access Single sign-on Planned
Personal data PII policies, per-chat and per-agent settings
The prompt Content policies on the prompt Planned
The answer Content policies on the answer Planned
Tools Tool rules, tool classes, tool modes Tool mode limits for everyone, a team, an agent or a person; content policies on tool arguments Planned
Grants Grants for a time, working alongside, Bypass; the installation's longest grant Approval limits for everyone, a team, an agent or a person; whom Bypass is for; alerts by mail
The record Decisions, without the text; the calls of a grant for auditors Agent actions for auditors, with an export as CloudEvents; policy names Planned

Content policies will refuse a prompt, an answer or a tool call by what it says.