Data processing agreement
Version in force as of 19 November 2026
Translation provided for information only. Only the French version of this document is legally binding.
Agreement concluded under Article 28 GDPR: it governs the personal data that Aurva Software processes in AIForeGuard for its customers.
1. Purpose and capacity of the parties
This data processing agreement (the “DPA”) is concluded under Article 28 of Regulation (EU) 2016/679 of 27 April 2016 (“GDPR”). It is appended to the terms of sale and forms an integral part of the contract concluded between the Publisher and the Customer, together with the Order Form and the terms of use.
The AIForeGuard service (the “Service”) is published by AURVA SOFTWARE, a simplified joint-stock company with a single shareholder (SASU) with share capital of €100, whose registered office is at 138 avenue de Versailles, 75016 Paris, France, registered with the Paris Trade and Companies Register under number 130 051 519 (the “Publisher”).
The Customer acts as controller for the Personal Data it has processed in the Service. The Publisher acts as processor, on behalf of and on the instructions of the Customer.
For the data it processes for its own purposes (demonstration requests and contacts, management of customer accounts and of the business relationship, invoicing, security of the Service, technical logs and error tracking), the Publisher acts as controller. That processing is described in the privacy policy and does not fall within the scope of this DPA.
2. Definitions
- Customer Data: the data collected, imported, entered or produced in the Service by the Customer, by its Users or by the components it deploys (connections to its platforms, browser extension, relay), including the resulting inventories, observations, alerts, decisions, files and reports.
- Personal Data: the personal data, within the meaning of Article 4 GDPR, contained in the Customer Data.
- Browser extension: the AIForeGuard module that the Customer may deploy on its employees’ browsers to detect and govern undeclared AI usage (“Shadow AI”).
- Live control: the feature that allows or blocks, at the time they are called, the tools of the Customer’s Microsoft Copilot Studio agents.
- Subprocessor: any provider that the Publisher uses to process Personal Data on behalf of the Customer.
- Connected Services: the third-party services that the Customer links to the Service under its own responsibility (section 9).
- Other capitalised terms, in particular Customer, User, Organisation, Customer Personnel and Order Form, have the meaning given to them in the terms of sale and the terms of use.
3. Description of the processing
Subject matter: the provision of the Service, an artificial-intelligence governance and security platform that enables the Customer to know, control and document the AI agents and AI usage within its organisation.
Nature of the operations: reading, with the permissions granted by the Customer, the configurations and metadata of its AI platforms (in particular Microsoft Copilot Studio, Microsoft 365 Copilot, Azure AI Foundry, AWS Bedrock, AgentCore and Quick, Google Gemini Enterprise, Mistral); receiving the browser extension’s observations; collection, normalisation, pseudonymisation by hashing, recording, comparison, consultation, export and deletion; producing alerts and sending them to the tools chosen by the Customer; deciding whether to allow or block a tool as part of live control.
Purposes: to enable the Customer, and the Customer alone:
- to inventory its AI agents and AI usage, and to detect changes to them;
- to detect undeclared AI usage (Shadow AI) through the browser extension and to apply its policies and browser rules to it;
- to define policies and receive alerts, in the Service or in its own tools;
- to conduct periodic reviews, to set up mission contracts for its agents and to produce audit files and reports;
- to control live, and block where appropriate, the tools of its Copilot Studio agents;
- to ensure the traceability of actions performed in the Service (audit log).
The Service uses no artificial-intelligence model and no language-model provider to process Customer Data. Connections to the Customer’s AI platforms are used solely to read their configurations and metadata; live control also receives, at each tool call, the items described in section 4, which are not retained.
Duration: the processing lasts as long as the contract, then for the time needed for the return and erasure operations provided for in section 15. The retention periods for each category of data are set out in section 4.
Location: Customer Data is stored in the European Union: database in Paris (France), server functions in the Paris region, secrets vault in France (France Central region). Some Subprocessors are established in the United States or operate global networks; these situations and the applicable safeguards are described in section 10 and on the Subprocessors page.
4. Data and data subjects
The Customer chooses the platforms it connects, the features it enables and the people covered by its deployment. Depending on those choices, the Service processes all or some of the categories below. Collection is based on lists of permitted fields: anything not on them is discarded.
Data subjects
- the Users of the Service designated by the Customer;
- Customer Personnel (employees, contractors and other staff) whose AI usage, devices or capacity as owner or author of an AI agent appear in the data collected;
- the people named in the configuration of the Customer’s AI agents.
Categories of data
- Users of the Service: identifier assigned by the identity provider (issuer and subject), email address, display name, role (administrator, security analyst, AI agent manager, auditor, reader), permissions and date of last sign-in; sessions, of which only a hash (SHA-256) of the token is kept; link with Slack (Slack identifier and name); mapping between Microsoft Entra groups and roles. For email-and-password accounts created by the Publisher, the password is kept only in hashed form (scrypt).
- AI agent inventory: configuration and components of the agents, limited to selected fields; agent instructions, kept only as a hash and a length; display names and work email addresses of the people named by an agent (owners, authors); number of conversations per day; Microsoft Entra sign-in logs of agent creators (device, operating system, compliance, application, dates), without IP address or location.
- Approval before publication: work email address of an agent’s creator, their answers to the approval request, the business owner’s opinion and the decision.
- Hold of unapproved agents (optional): address of the flow installed in the Customer’s tenant, or sign-in name of the linked administrator and their delegated token (secret vault); the Publisher performs only two actions on the Customer’s platform: putting an unapproved agent in quarantine, lifting the quarantine on approval.
- Microsoft 365 Copilot audit logs: metadata only (application, agent, extensions, sensitivity label identifiers, jailbreak detections, model provider); user identifier, kept as a hash computed with a key specific to the Customer’s Organisation; administrative actions on agents (author, action, date); number of licences.
- Microsoft Defender: device name, Microsoft Entra device identifier, local account name, host or command of the detected MCP server (AI assistants installed on workstations).
- Live control: for each decision, the agent, the tool, the parameter names, the outcome, the Microsoft Entra user reference and the conversation reference. The user’s message, the conversation history and the values passed to the tools are read in memory to make the decision and are not retained.
- Browser extension: origin host and destination host among AI services, type of observation, signal codes, daily counters per host (visits, messages, requests, blocks); email address of the browser profile, only as a hash (HMAC); directory attachment (department or division, groups), with identifiers kept as hashes, without any name or email in clear text; access requests from an employee (name, work email and reason entered by the employee); employee records (name and optional reference). If the Customer turns on per-person detail (“All usage per person”), the name, size and SHA-256 fingerprint of each file whose upload to an AI service was blocked, with the person and time, as evidence; never the file content, and these items are erased at the end of the chosen period (30, 90 or 180 days) or as soon as the Customer turns the option off. Document labels are read locally, on the device. In Netskope mode, the content of a file passes through the AIForeGuard server to the Customer’s Netskope tenant, without being retained. When the Customer installs the optional relay, per-browser detail remains with the Customer (8 days) and only totals per unit are sent to the Service.
- Google Workspace directory (optional): read-only access, authorised by one of the Customer’s Google administrators (
admin.directory.user.readonly,admin.directory.orgunit.readonly,admin.directory.group.readonly,openid email), once an hour: department, organisational units and groups. Only hashes and the attachment (department, group) are kept, never the name or email in clear text. The refresh token is kept in the secrets vault; unlinking the directory revokes this token with Google. - Emails sent from the Customer’s mail (optional): through Microsoft 365 (sending mailbox designated by the Customer) or Gmail (
gmail.sendon a dedicated mailbox, token in the secret vault); no email is read or kept. - Alerts sent to the Customer’s tools (Microsoft Teams, Slack, Jira, ServiceNow, Microsoft Sentinel, webhook): title, priority, reason, agent name, platform and connection, dates, link to AIForeGuard; in Slack, the name of the person who takes charge of the alert and of the person who resolves it. Alerts never contain document content.
- Governance and audit log: decisions, justifications and comments entered by Users (reviews, mission contracts, files, reports); log of who did what and when in the Service, append-only and unalterable.
What the Service never retains
- the text of conversations with AI agents, or the prompts and responses of Microsoft 365 Copilot;
- agents’ files and the values of their environment variables;
- with the browser extension: the prompts, responses and files sent to AI services, cookies, page titles, URL paths and URL parameters;
- the IP address and location contained in the sign-in logs of agent creators;
- document content, in the alerts sent to the Customer’s tools.
Sensitive data
The Service is not intended to process special categories of data within the meaning of Article 9 GDPR, or data relating to criminal convictions and offences (Article 10 GDPR). The Customer undertakes not to configure the Service to collect such data, in particular through browser rules, free-text fields (reasons, comments, labels) or the connections it sets up. If it does so, it alone is responsible for the lawfulness of that processing.
Retention periods
| Data | Retention period |
|---|---|
| Inventory and current state of agents, settings, and what is still in force: open cases, last decided review of each agent, decisions and exceptions not yet expired, refused publication requests, notices sent, mission contract in force, with the evidence they cite; reports produced | Term of the contract, unless a shorter period is agreed with the Customer |
| Display names of the people named by agents | 90 days after the last reading |
| Live control decisions | 90 days |
| Audit log, agent events, notifications, browser extension events, activity counters, AI usage, administration events, collection history, observed agent states, evidence items and detected changes that no file cites any more, closed cases (reviews, requests, resolved alerts, plans), superseded versions and declarations | 1, 2 or 3 years, as chosen by the Customer in the Service (3 years by default) |
| Exports (files) and invitations | 7 days |
| Sessions | Deleted 7 days after they expire |
| Database backups | 7 days (rotation) |
| Per-browser detail in the relay installed at the Customer’s premises (optional) | 8 days, at the Customer’s premises |
| Content read by live control; file passed on in Netskope mode | Not retained (read in memory or merely in transit) |
At the end of the contract, Customer Data is erased under the conditions of section 15, even if its retention period has not expired.
5. Customer instructions
The Publisher processes Personal Data only on documented instructions from the Customer, including with regard to transfers to a third country. The contract, this DPA and the settings chosen in the Service by the Customer’s administrators (connections and permissions granted, deployment and rules of the browser extension, live control, alert destinations, User roles) constitute those instructions.
The permissions granted by the Customer define the limits of the Publisher’s access. For Microsoft, these are the read permissions AgentIdentity.Read.All, AuditLog.Read.All, GroupMember.Read.All, User.Read.All, ThreatHunting.Read.All, LicenseAssignment.Read.All and ActivityFeed.Read, delegated access to Dynamics and Azure Resource Manager, and the “AI Reader” role. The Publisher uses them only for the purposes set out in section 3.
The Publisher does not use Customer Data for its own purposes, for statistical, commercial or advertising purposes, or to train artificial-intelligence models. Only the figures needed to manage the contract (for example the number of AI agents monitored) are visible to the Publisher. It does not sell Customer Data and does not disclose it to any third party other than the Subprocessors and the tools designated by the Customer.
Data obtained from Google APIs. When the Customer connects its Google Workspace directory, the use of data received from Google APIs complies with the Google API Services User Data Policy, including its Limited Use requirements: this data is used only for the features of the Service; it is not used for advertising, is not transferred to any third party other than the Subprocessors needed to operate the Service, is not used to train artificial-intelligence models and is not read by any person, except for security purposes, to comply with a legal obligation or at the Customer’s request.
The Publisher immediately informs the Customer if, in its opinion, an instruction infringes the GDPR or another provision of Union or Member State data protection law.
Where Union or Member State law to which it is subject requires it to process Personal Data other than on the Customer’s instructions, in particular to transfer it, the Publisher informs the Customer before the processing, unless that law prohibits such information on important grounds of public interest.
6. Customer obligations
As controller, the Customer determines the purposes and means of its processing and ensures that it complies with the law. In particular, the Customer undertakes to:
- have a legal basis for each processing operation it carries out using the Service (Article 6 GDPR);
- inform its employees and other data subjects in advance, clearly and fully, of the processing carried out, in particular before deploying the browser extension, live control or any other means of monitoring their activity (Articles 12 to 14 GDPR and, in France, Article L1222-4 of the French Labour Code);
- inform and consult employee representatives where the law so requires, before deciding to implement such means (in France, the social and economic committee, Article L2312-38 of the French Labour Code), and comply with the equivalent rules of the other countries where it deploys them;
- ensure that its deployment is proportionate, by limiting the features enabled, the browser rules, the people covered and its Users’ rights to what is necessary for specified and legitimate purposes;
- carry out a data protection impact assessment (DPIA) where the processing is likely to result in a high risk to the rights and freedoms of individuals, in particular where it allows systematic monitoring of employees’ activity (Article 35 GDPR);
- not configure the Service to collect sensitive data (section 4);
- grant access permissions to its platforms and tools (Microsoft, Amazon Web Services, Google, Mistral, directories, alerting tools, Netskope) in compliance with its own contracts with their providers, and withdraw them when it no longer wishes them to be used;
- designate its Users, assign them appropriate roles and withdraw their access when they no longer need it;
- respond to requests from data subjects, with the assistance provided for in section 11.
Processing carried out in the Customer’s tools after an alert is received (Microsoft Teams, Slack, Jira, ServiceNow, Microsoft Sentinel, webhook) is the Customer’s sole responsibility and is governed by the contracts between the Customer and the providers of those tools.
7. Confidentiality
The Publisher ensures that the persons authorised to process Personal Data have committed themselves to confidentiality or are under an appropriate statutory obligation of confidentiality.
Access to Customer Data is limited to the persons who need it to operate, maintain or troubleshoot the Service. The Publisher’s administration console gives no access to the Customer’s governance data (agents, files, connections, evidence); every action the Publisher performs there for an Organisation is recorded in that Organisation’s audit log.
8. Security
The Publisher implements the appropriate technical and organisational measures required by Article 32 GDPR, taking into account the state of the art, the costs of implementation, the nature, scope, context and purposes of the processing, and the risks to the rights and freedoms of individuals. These measures are described in the appendix (section 17).
These measures may evolve with the state of the art. The Publisher will not lower them below the level of protection in place on the date the contract comes into force.
The Customer remains responsible for the security of the elements under its control: its identity provider and the authentication of its Users (in particular multi-factor authentication), the devices and browsers on which it deploys the extension, the relay it installs where applicable, and the tools to which it sends alerts.
9. Subprocessors
The Customer gives the Publisher a general authorisation to use Subprocessors. The up-to-date list, with their purpose, the data concerned, their location and the applicable safeguards, is published on the Subprocessors page.
The Publisher informs the Customer of any intended addition or replacement of a Subprocessor at least one (1) month before it is brought into service. The Customer may object within that period, on reasonable grounds relating to data protection; failing agreement between the parties, the Customer may terminate the part of the Service concerned, without penalty.
The Publisher contractually imposes on each Subprocessor data protection obligations equivalent to those of this DPA, in particular sufficient guarantees to implement appropriate technical and organisational measures. It remains fully liable to the Customer for the performance of their obligations.
Connected Services (for example Microsoft Entra ID, Microsoft 365, Amazon Web Services, Google Workspace, Slack, Jira, ServiceNow or Netskope) are not Subprocessors of the Publisher: the Customer chooses them, links them to the Service under its own responsibility and remains bound to their providers by its own contract. They are listed, for information, on the Subprocessors page. The same applies to the notification service built into the browser the Customer deploys (for example Google for Chrome, Microsoft for Edge), through which the Service wakes the extension with an empty message containing no content.
10. Transfers outside the European Union
Customer Data is stored in the European Union (section 3). However, some Subprocessors are established in the United States, may access the data from a third country for operations and support, or operate global networks: delivery of the Service’s static files, delivery of the browser extension’s rule files and the Publisher’s email service.
These transfers are governed by the safeguards provided for in Chapter V GDPR: an adequacy decision where applicable, in particular where the provider participates in the EU–US Data Privacy Framework, and failing that the European Commission’s standard contractual clauses. They are accompanied by appropriate supplementary measures: encryption in transit, signing and encryption of the extension’s rule files (which the delivery network cannot read) and pseudonymisation of identifiers. Details are given on the Subprocessors page.
11. Data subject rights
The Service provides the Customer with features that help it respond to data subject requests itself: consultation of the data, management of Users and their roles, export of the audit log, review files and reports.
If a data subject contacts the Publisher directly, the Publisher does not respond on the merits, unless instructed by the Customer, and forwards the request to the Customer without undue delay.
Where the features of the Service are not sufficient, the Publisher provides the Customer with reasonable assistance, taking into account the nature of the processing and the information available to it.
12. Assistance to the Customer
The Publisher assists the Customer, to a reasonable extent and taking into account the nature of the processing and the information available to it, in complying with the obligations of Articles 32 to 36 GDPR: security of processing, notification and communication of personal data breaches, data protection impact assessment and prior consultation of the supervisory authority.
For its impact assessment and for informing its employees and their representatives, the Customer may obtain on request, at privacy@aiforeguard.com, a description of the data processed by each feature (in particular the browser extension and live control), of their retention periods and of the security measures, for which this DPA serves as the basis.
13. Personal data breaches
The Publisher notifies the Customer of any personal data breach affecting Customer Data without undue delay and no later than forty-eight (48) hours after becoming aware of it. The notification is sent to the Customer’s administrators or to the contact it has designated.
The notification describes the nature of the breach, the categories and approximate number of data subjects and records concerned, the likely consequences and the measures taken or proposed to address the breach and mitigate its effects. Information that is not yet available is provided as it becomes available.
The Publisher promptly takes reasonable measures to contain the breach and limit its consequences, and documents the facts, their effects and the measures taken.
Notification to the supervisory authority and, where applicable, communication to data subjects are the Customer’s responsibility, as controller. The Customer may report any suspected security incident to the Publisher at privacy@aiforeguard.com.
14. Audits and documentation
The Publisher makes available to the Customer the information necessary to demonstrate compliance with the obligations of Article 28 GDPR.
The Customer may have one audit carried out per year, at its own expense, either itself or through an independent auditor bound by a confidentiality obligation who is not a competitor of the Publisher, with thirty (30) days’ notice. The audit takes place during business hours, without disrupting the Service or exposing other customers’ data.
The Publisher may meet this obligation by providing the documentation of its security measures and, where available, the audit reports or certifications of its Subprocessors.
15. Data at the end of the contract
Throughout the term of the contract, the Customer may export its data using the features of the Service: audit log, review files and reports (PDF). It is the Customer’s responsibility to make these exports before the contract ends; export files are kept for 7 days, within which the Customer downloads them.
During the thirty (30) days following the end of the contract, for whatever reason, the Publisher helps the Customer, on request, to obtain any exports it has not made. When that period expires, or earlier at the Customer’s written request, the Publisher carries out the complete erasure of the Customer’s Organisation, which deletes Customer Data from its active systems, including the connections and the associated secrets. This operation is carried through to completion, even in the event of a technical interruption, and gives rise to a receipt containing no personal data, provided to the Customer on request.
Copies held in the database backups disappear at the end of their rotation cycle, which is seven (7) days. Versions of secrets kept in the vault are permanently purged ninety (90) days after their deletion, a period that protects against accidental erasure. Until then, these copies remain protected by the measures of this DPA.
The Publisher retains nothing beyond this except what Union or Member State law requires it to retain, and only for the period required.
Components installed at the Customer’s premises (browser extension, relay) and the permissions it has granted in its platforms remain under the Customer’s control: it is for the Customer to uninstall or withdraw them.
16. Term, liability and amendments
This DPA takes effect on the date the contract comes into force and remains in force for as long as the Publisher processes Personal Data on behalf of the Customer.
In the event of any conflict between this DPA and the terms of sale or the terms of use, this DPA prevails with regard to the processing of Personal Data, unless the Order Form contains a provision that is more protective of that data.
Each party is liable for the obligations that the GDPR places on it. Between the parties, each party’s liability under this DPA is subject to the conditions and limits set out in the terms of sale.
The Publisher may amend this DPA to reflect changes in the law or in the Service. Any substantial amendment is brought to the Customer’s attention at least thirty (30) days before it comes into force.
This DPA is governed by French law. Any dispute relating to its performance is handled in accordance with the dispute resolution clause of the terms of sale.
Questions about this agreement: privacy@aiforeguard.com.
17. Appendix — Technical and organisational security measures
This appendix describes the measures implemented by the Publisher on the date this DPA comes into force, under Article 32 GDPR.
Separation of Organisations and access control
- Isolation of each Organisation by row-level security (RLS) enforced in the database; separate database roles, with no right to bypass this security; isolation checked after every change to the schema.
- Roles and permissions assigned to each member (administrator, security analyst, AI agent manager, auditor, reader), with optional mapping between Microsoft Entra groups and roles.
- Sign-in through the Customer’s identity provider; short sessions (8 hours at most, closed after 2 hours of inactivity), of which only a hash of the token is kept; passwords of accounts created by the Publisher kept only in hashed form (scrypt).
- Dual approval (“four eyes”) for agent mission contracts, lifecycle stage changes and regulatory profiles.
Secrets and encryption
- Secrets (keys, tokens and integration passwords entrusted by the Customer) kept only in a Microsoft Azure Key Vault located in France, never in the database; access to the vault through identity federation, with no stored secret.
- Encryption in transit (TLS).
- Browser rules signed (Ed25519) and encrypted (AES-GCM): the delivery network can neither read nor alter them, and the extension checks their signature before applying them.
Minimisation and pseudonymisation
- Collection limited to lists of permitted fields; agent instructions kept only as a hash and a length.
- Identifiers pseudonymised by hashing: Microsoft 365 Copilot users (key specific to the Organisation), browser profile email address (HMAC), directory identifiers.
- No retention of the content of conversations, of prompts and responses to AI services, or of files; in-memory reading only for live control.
- Deletion of data when the retention periods of section 4 expire.
Application protection
- Protection against cross-site request forgery (CSRF), rate limits, strict content security policy.
- Allow-lists of hosts for outbound calls.
- Technical error tracking without personal data: IP address, location, user, cookies, headers and content are filtered out before sending.
Traceability
- Append-only, unalterable audit log: who did what and when in the Service.
- Actions taken by the Publisher from its administration console recorded in the audit log of the Organisation concerned; no Customer governance data accessible from that console.
Continuity and erasure
- Daily database backups, kept for 7 days.
- Complete erasure of an Organisation at the end of the contract, carried through to completion even in the event of an interruption, with a receipt containing no personal data; permanent purge of secret versions 90 days after their deletion.
Organisation
- Access by the Publisher’s personnel limited to what is necessary to operate, maintain and troubleshoot the Service, under a confidentiality commitment.
- Subprocessors contractually bound by data protection obligations (section 9).
- Procedure for notifying personal data breaches to the Customer (section 13).