Data Security Kit in the Face of AI

Free resource in French · Word + Excel Kit

Data Security Kit in the Face of AI

Is your AI project ready for launch? In two hours, identify its risks, mitigate them, and make an informed decision: go, go with conditions, or no go. Fourteen risks explained in client-centric language, eight red lines, a decision that's signed. No jargon, no fear-mongering.

📄 Word, 65 pages📊 Excel workbook, 10 tabs🇫🇷 Up-to-date with ANSSI, OWASP and IBM 2026 standards🆓 Free
IAR1R3R5R111Realizefourteen risk sheets2Evaluateprofile and rating3Mitigatemeasures and questions4Decidego, under certain conditions, no goGO

The AI ​​project sponsor's security and data kit, in brief

What it is
A decision-making method for the person leading an AI project: become aware of the fourteen security and data risks specific to AI projects, including those of AI in action, assess the exposure of their project, choose the measures that reduce them, then decide go, go conditionally or no go, with a decision sheet to sign and a plan to be ready on the day of an incident.
For who
The sponsor of an AI project, not the CISO: business directors, sponsors, AI project managers, transformation directors, in groups, banks, insurers, public institutions and industrial companies; and their interlocutors CISO, DPO, CIO and purchasing, who will find the questions that will be asked of them.
Format
65-page Word document (the four-step process, the post-go phase and six appendices) and Excel workbook in 10 tabs: instructions for use, profile, risks, rating, mitigation, red lines, decision, incidents, supplier, agents.
What it contains
The fourteen risk sheets (leaks via prompts, over-sharing, shadow AI, prompt injection, agents exceeding their mandate, generated code and applications, AI supply chain, vendor and data, impactful hallucination, fraud and deepfakes, invisible incidents, non-compliance, agent identities and proliferation, memory and poisoned sources); what changes when AI acts (lethal triad, rule of two, autonomy levels A0 to A4); the project profile in twelve questions and four exposure levels; the severity × likelihood rating and its heat map; 85 measures per risk, the ten rules of generated code, the five rules of vibe coding, data preparation before Microsoft 365 Copilot (often called "Copilot Premium"); ten questions for the CISO, ten for the DPO, thirty for the vendor; the eight red lines, the decision grid, the form to be signed; Seven incident response sheets, the notification tree (GDPR, NIS2, DORA, AI Act, Cyber ​​Resilience Act), and the tabletop exercise. Appendices include: the memo to the management committee, the thirty-six-point checklist, fourteen frequently asked questions, a glossary of thirty-seven terms, and dated sources.
What is it based on?
ANSSI publications on AI security (recommendations for a generative AI system from April 2024, CERT-FR threat summary from February 2026, zero trust principles published with the BSI in August 2025); the CERT-FR alert on desktop agents (April 13, 2026) and the Five Eyes agencies' guide on the adoption of agentic AI (May 1, 2026); the OWASP GenAI Security Project repositories (Top 10 LLM 2026, Top 10 agentic applications, MCP Top 10), MITRE ATLAS and the NIST agent initiative; the CoSAI incident response framework; the IBM Cost of a Data Breach 2026 report and the Veracode 2026 report on the security of generated code; the CNIL's recommendations on the deployment of generative AI and Microsoft's blueprint on over-sharing.
Price
Free, in exchange for a professional email address. No unsolicited emails.
Update
Version 2, updated October 26, 2026: AI acting at the center (lethal triad, autonomy levels, two new risks R13 and R14, two new red lines L7 and L8, "9 Agents" tab), five more real incidents, DORA and the Cyber ​​Resilience Act in the notification tree, supplier questionnaire increased to thirty questions. Reviewed every six months, and at each milestone: French law transposing NIS2 (known as the "Resilience" law, withdrawn from the National Assembly's agenda on October 6, 2026 and not yet promulgated to our knowledge as of October 26, 2026), December 2, 2027 (high-risk obligations of Annex III of the AI ​​regulation).
What it doesn't replace
The risk analysis of your CISO (of the EBIOS RM type), the opinion of your DPO or a legal opinion: the kit informs the decision of the sponsor, it does not delegate it and does not replace it.
Go Further
sponsor's GDPR kit, AI Act kit from the sponsor, AI and Social Dialogue Kit, frugal AI kit, responsible AI kit, AI Systems Governance Kit.

The scene repeats itself in every large organization. The pilot project was a success, users are clamoring for more, and the committee wants a deployment date. Then the CISO asks three questions: what data will the tool be able to read, what will the agent do on its own, and what will happen if someone hijacks it? No one around the table has the answer, and the project is back to square one for another quarter of analysis, the purpose and outcome of which the sponsor doesn't understand. Conversely, other projects go ahead without anyone asking these questions, and that's often where incidents originate.

This kit provides the client with the tools to fulfill their role without becoming a cybersecurity expert: a four-step process—awareness, assessment, mitigation, and decision—that takes place over two hours with the project manager and culminates in a written decision. It does not replace the work of the CISO; it allows them to present the project to the CISO with a pre-defined project, the right questions, and clearly assigned responsibility where it belongs: with the project leader.

1. AI has changed the threat, and your project falls within its scope.

Attackers are using AI, and AI systems have themselves become targets, and then targets that act. In February 2026, CERT-FR observed that generative AI is integrated into all stages of the attack chain (social engineering and deepfakes, malicious code, analysis of stolen data) and that AI systems are subject to attacks specific to them: poisoning of training data, backdoors in models, prompt injection. The facts below are those on which the kit is based, along with their source and date.

DateFact or figureWhat this changes for a sponsor
April 2023Samsung: Engineers paste source code and internal meeting minutes into ChatGPT; the company then restricts the use of public generative AI.Shadow AI is born from a legitimate unmet need: banning is not enough, a validated alternative must be offered.
June 2025EchoLeak (CVE-2025-32711), discovered by Aim Security: an email containing a hidden instruction was enough for Microsoft 365 Copilot to exfiltrate internal data without any user action. Microsoft patched it server-side before publication.Any content that the AI ​​reads (email, document, web page) is a potentially hostile input.
2025USENIX Security, Spracklen et al.: out of 576,000 generated code snippets, approximately 19,7% of the recommended software packages did not exist; attackers then publish malicious packages under these names (slopsquatting).Each component that the AI ​​proposes to install must be checked before use.
July to September 2025Agents and agent tools caught in a trap: Replit agent deletes production database despite freeze instruction (July); hidden comment makes Comet agentic browser interact with user session (Brave, August); ForcedLeak hijacks Salesforce Agentforce agent via public web form (Noma Security, September); first malicious MCP server discovered (Koi Security, September).A written instruction in a prompt is not a control: only technical rights and human validation are.
November 2025 (?)Anthropic describes the GTG-1002 espionage campaign: a hijacked code assistant, linked to tools via MCP, executes 80 to 90% of tactical operations against some thirty organizations.AI is no longer just the attacker's tool: it is a target that acts.
4th February 2026ANSSI and CERT-FR, summary of the 2025 threat to generative AI: central recommendation to strictly compartmentalize AI systems and sensitive environments.The compartmentalization is decided at the project scoping stage, not after deployment.
February 2026Snyk, ToxicSkills: out of 3,984 agent "skills" published on two marketplaces, 13,4% had a critical vulnerability and 76 contained confirmed malicious code.An agent extension inherits all its rights: it installs like software, with a review and catalog.
April 13 and May 1, 2026CERT-FR advises against deploying desktop agents that control the workstation and browser on workstations, except in an isolated test environment; the Five Eyes network agencies publish their first joint guide on the cautious adoption of agentic AI.Inventory of agents, specific identities, fair rights, human validation of high-impact actions, gradual deployment.
Summer 2026IBM, Cost of a Data Breach Report 2026: global average cost of a breach of $4,99M (+12%); more than one in four organizations has suffered an attack using AI; 43% of incidents involve shadow AI, for an average cost of $5,39M compared to $4,63M; 92% of organizations affected by an AI-related incident did not have access control over their AI tools.Costly incidents primarily stem from a lack of rules and access rights, not from exotic attacks.
28th July 2026 Veracode, 2026 GenAI Code Security Report: AI-generated code passes security tests in only 56% of cases, 44% of tasks introduce a vulnerability, and the success rate drops to 15% for cross-site scripting. "The syntax is solved, but the security is not."No generated code goes into production without human review or automated analysis.
August and September 2026OWASP GenAI Security Project: Top 10 for LLM Applications 2026 (published August 3, 2026), Top 10 for Agentic Applications 2026 (published December 2025), MCP Top 10 (beta), Agent Control Standard (version 0.1, September 2026).Agents and connectors now have their own risk lists: these are yours as soon as the project takes action.

The kit also specifies the deadlines that apply on the day of an incident (72 hours to notify the CNIL of a personal data breach, 24 hours for early warning of an entity subject to NIS2 once the French law is enacted, 4 hours after classification for a major incident of a financial entity subject to DORA, 24 hours to report an actively exploited vulnerability under the Cyber ​​Resilience Act since September 11, 2026, 15 days for a serious incident on a high-risk system as defined by the AI ​​regulation, from December 2, 2027) and the timetable of the "Omnibus AI" regulation (Regulation (EU) 2026/1744), which postponed the high-risk obligations to December 2, 2027 for Annex III and to August 2, 2028 for Annex I, reformulated Article 4 (control of AI by staff) into an obligation of means, without deferring the application, and only granted a deadline until December 2, 2026 for the marking of content provided for in Article 50 (transparency).

2. The four-step process, from awareness to decision

The kit follows a single thread, from assessment to signature. Each step has its own chapter in the Word document and its own tabs in the Excel workbook, which calculates the level of exposure, the relevance of the risks, the scores, and the decision. The yellow cells are to be filled in; the gray ones are pre-calculated.

StageWhat we doWhat comes out of it
1. Become awareRead the fourteen risk fact sheets: the risk in one sentence, how it arises, a thought-provoking case, warning signs in your project, the question to ask yourself, and who should answer it.A common vocabulary with the CISO and the DPO (tab "2 Risks")
2. EvaluateAnswer the twelve questions in the project profile, filter the relevant risks, and rate each one based on severity × likelihood.An exposure level from 1 to 4 and a heat map (tabs "1 Profile" and "3 Rating")
3. MitigateChoose risk-based measures with a designated person responsible, a defined effort, and a deadline; ask questions of the CISO, DPO, and vendorA plan of measures and residual risks (tabs "4 Mitigation", "8 Supplier" and "9 Agents")
4. DecideCheck the eight red lines, apply the decision grid, and ensure you are prepared for the incident.Go, conditional go or no go, and a decision sheet to sign (tabs "5 Red Lines", "6 Decision" and "7 Incidents")

3. The fourteen risks a sponsor should know

Each risk is described on a one-page sheet, written for a decision-maker, not a specialist. Not all risks apply to every project: the project profile indicates which ones are relevant to yours. Here are the fourteen risks, along with the question the toolkit asks you to consider; R13 and R14, new in version 2, cover AI in action.

RisqueIn one sentenceThe question to ask yourself
R1. Data leakage via promptsEmployees paste confidential data (contracts, code, customer data, personal data) into an AI tool which stores or reuses it.What data will go into the tool, and where will it go?
R2. OversharingAI finds and summarizes documents that the user technically has access to but should not see: poorly protected HR files, salaries, reorganization plans, sites open to "everyone".Are our access rights secure before we plug in an AI that reads everything?
R3. Shadow AI around the projectUnvalidated AI tools or agents are being used in parallel, or the project itself is slipping out of the control of the IT department.What are my teams already using, and why?
R4. Prompt injectionA hidden instruction in an email, document, web page or ticket can hijack AI: exfiltration, unwanted actions, disinformation.What content that the AI ​​will read is not under our control?
R5. Agent exceeding their mandateAn agent performs irreversible actions (send, delete, pay, modify) without validation, or chains tools beyond the intended scope.What is the agent authorized to do, and who presses the button?
R6. Generated Code and ApplicationsCode produced by an assistant, or an entire application created in natural language by a professional ("vibe coding"), goes into production without review or analysis.Who reviews, who tests, and how do we know if a code comes from AI?
R7. AI Supply ChainUncontrolled software packages, MCP servers and third-party connectors, agent skills and plugins, downloaded models, AI browser extensions enter the project without control.What third-party components are involved in the project, and who has verified them?
R8. Supplier and dataData reuse for model training, non-EU location, prompt retention, subcontractors, reversibility, dependency.Do I have the contract and written responses from the supplier on these five points?
R9. Impact HallucinationA false but plausible answer is used for a customer, legal, financial, medical, or HR decision without human verification.Where would an AI error really hurt, and who checks before acting?
R10. Fraud and deepfakesImpersonation of a manager, supplier or customer by a generated voice or video, personalized phishing.Can our validation procedures (payment, change of bank details, access) withstand a perfect voice or video?
R11. Invisible IncidentNo logging of prompts, responses and actions: no one would know that a leak or diversion had taken place, nor could they prove it.If an incident were to happen tomorrow, what would we see, and who would be notified?
R12. Non-conformityPersonal data without legal basis or information, obligations of the AI ​​regulation, regulated sectors, social dialogue forgotten.Have my DPO and my lawyer seen the project?
R13. Identities and Agent ProliferationAgents multiply without inventory, without owner and without their own identity: they operate with an employee's account or a shared account, keep tokens for months and continue operating after their creator leaves.Who are our agents, who is responsible for them, and what rights do they have to act?
R14. Memory and poisoned sourcesA document trapped in the document database, a modified page on the intranet, or false information stored in an agent's memory permanently influences the responses and actions of the AI ​​for all users.Who can write in what AI reads and remembers?

4. Assess and reduce: profile, rating, measures and questions to ask

The assessment starts with the project itself, not a list of threats. Twelve questions describe its profile: the data processed, the users, what the AI ​​does (read, write, recommend, act independently), the hosting, the content it reads, the code, third-party components, the impact of an error, logging, governance, the "lethal triad" (a single agent accessing private data, reading uncontrolled content, and communicating externally), and who creates the agents and applications. Each answer is worth 0 to 3 points; the total, out of 36, gives an exposure level from 1 (low) to 4 (critical) and indicates which risks are relevant. Each risk is then rated by severity and likelihood on a scale of 1 to 4, based on default values ​​that you adjust, and the spreadsheet places all fourteen on a heat map.

To reduce risks, the kit offers 85 ready-to-assign measures, each with a designated person responsible, effort level, and deadline. For an agent, for example: a level of autonomy (from A0, read, to A4, act alone) defined in their mandate, a unique identity and designated owner, reduced privileges, human validation before any irreversible action, a whitelist of tools and actions, a sandbox for testing, and a log of each action with an emergency stop. Three thematic sheets go further: the ten rules of AI-generated code, the five rules of vibe coding for applications created by business users, and data preparation before Microsoft 365 Copilot (often called "Copilot Premium"), which queries the entire company graph unlike Copilot Chat, in eight points (inventory of over-sharing, Restricted Content Discovery on sensitive sites, sensitivity labels, DLP rules, cleaning up links open to "everyone", Purview DSPM for AI, pilot of 50 people on a specific scope, review of reports before extension).

The sponsor doesn't have to find the answers alone: ​​they must ask the right questions to the right people. The kit provides ten questions for the CISO, ten for the DPO, and thirty for the vendor (including agent identities, output domains, skills and connectors, component nomenclature, NIS2, DORA, and the Cyber ​​Resilience Act), rated out of 60 in the "Vendor 8" tab: from 48 to 60, the vendor is solid; from 30 to 47, clauses need to be negotiated before signing; below 30, it's a red line to discuss with the CISO.

5. Decide: eight red lines and a go grid, conditional go, no go

The decision is not based on an average of scores. Eight red lines apply to all projects, regardless of size: failing to meet even one is enough to result in a "no go" decision. If the project has no agents, lines 7 and 8 are automatically met.

Red lineWhat must be true before the go
L1. DataNo secret, classified or sensitive data as defined by the GDPR is sent to a service that has not been qualified by the CISO and the DPO.
L2. AgentsNo agent can perform an irreversible action (external sending, deletion, payment, modification of reference data) without human validation.
L3. SupplierThe supplier has committed in writing not to reuse our data to train its models, or the opt-out option is enabled and verified.
L4. LoggingPrompts, responses, and actions are logged, and someone is designated to review them.
L5. IncidentAn incident response plan exists for this project: quick reference guides, contacts, notification tree.
L6. Critical RisksEvery critical risk after mitigation has a designated person responsible who accepts it in writing.
L7. Lethal TriadNo agent can combine, without human validation of its sorties, the three capabilities of the lethal triad: access to private data, reading of uncontrolled content, communication to the outside or action on systems.
L8. Register of agentsEach agent in production is registered, has its own identity (neither personal nor shared account), a named owner and a tested emergency stop.
DecisionWhenWhat follows
No goAt least one red line is not lifted, or a critical residual risk is not accepted.Reduce the scope (less data, fewer users, read-only instead of action) or change the solution (enterprise version, EU hosting, different provider), then repeat the process.
Go under certain conditionsAll red lines are lifted, but there remains at least one high or critical residual risk that must be accepted, or the project involves an agent acting alone and passing the test for the first time.The conditions are the remaining dated measures, with a review milestone at most 90 days (30 days for an agent acting alone).
GoAll red lines have been lifted, no residual risk exceeds the moderate level and no agent acts alone without having already passed a review.90-day review, six indicators to monitor, reassessment with each extension of the scope.

The decision is formalized on a one-page document: exposure level, number of risks per level before and after mitigation, status of red lines, decision, conditions and deadlines, residual risks accepted and by whom, and review date. The sponsor signs it; the CISO and the DPO add their opinions.

6. Be prepared for the incident: the condition for going

An AI project isn't ready if it doesn't know what to do when something goes wrong. The kit provides seven half-page quick reference guides, built on the CoSAI AI-adapted incident response cycle (detect, contain, notify, inform, learn): data leak via a prompt, prompt injection or AI hijacking, an agent acting outside its mandate, vulnerable AI-generated code discovered in production, a hallucination leading to a harmful decision, deepfake fraud or impersonation, and a compromised agent or agent component. It also includes a one-hour pre-deployment tabletop exercise with six participants (sponsor, CISO, DPO, CIO, business stakeholders, and communications representatives), five questions to answer, and a one-page report.

SituationWho notifies whomTime limit
Personal data breachThe DPO notifies the CNIL (GDPR, Article 33), and the data subjects if the risk is high.72 hours after becoming aware of it
Significant incident in an entity subject to NIS2 (once the French law is enacted)The entity informs ANSSIEarly warning within 24 hours, notification within 72 hours, final report within one month
Major ICT-related incident at a financial entity (DORA)The entity informs the ACPR or the AMF4 hours after classification as major and no later than 24 hours after detection; interim report within 72 hours; final report within one month
Vulnerability actively exploited in a digital product you publish (Cyber ​​Resilience Act, since September 11, 2026)The product manager informs the designated CSIRT and ENISAAlert within 24 hours, notification within 72 hours, final report within 14 days after the fix
Serious incident on a high-risk system (AI regulation, article 73, applicable from 2 December 2027)The system provider informs the market surveillance authority15 days; 10 days in case of death; 2 days for a critical infrastructure disruption or a large-scale breach
In all casesSupplier, cyber insurer, management, communication; filing a complaint in case of fraudAccording to the quick reference guides

7. What the kit contains

The Word document has 65 pages: the user manual, with what the kit does not replace; why this kit, nine real-world cases and what changes when AI acts (lethal triad, rule of two, autonomy levels A0 to A4); step 1, the fourteen risk sheets; step 2, the project profile in twelve questions, the filter of relevant risks and the severity × likelihood rating; step 3, the measures per risk, the ten rules of generated code, the five rules of vibe coding, agents, MCP servers, skills and agentic browsers, data preparation before Microsoft 365 Copilot, the ten questions for the CISO, the ten questions for the DPO and the thirty questions for the supplier; step 4, the eight red lines, the decision grid, the decision sheet to be signed and the "incident ready" condition (seven quick reference sheets, notification tree, tabletop exercise); The post-go phase (90-day review, six indicators, reassessment triggers). Six annexes complete the guide: the note to the management committee ready to adapt, the thirty-six-point checklist, fourteen frequently asked questions, the glossary of thirty-seven terms, dated sources and a page for further reading.

The Excel workbook "Tools in the Kit" has ten tabs, in the order of the process: "0 Instructions" (the process, the color key, the warning, what's changing in version 2); "1 Profile" (the twelve questions in drop-down lists and the calculated exposure level); "2 Risks" (the fourteen risks and their calculated relevance); "3 Rating" (severity, likelihood, score, level and 4 × 4 heat map); "4 Mitigation" (85 pre-filled measures with responsible party, effort, deadline, status, and the residual per risk); "5 Red Lines" (the eight, with cross-checks with the agents' tab); "6 Decision" (the calculated decision with its reason, then the form to be completed); "7 Incidents" (the register of incidents and notifications, including DORA and the Cyber ​​Resilience Act); "8 Supplier" (the thirty questions scored out of 60 and their interpretation); "9 Agents" (the register of twenty agents with level of autonomy, lethal triad, identity, owner and emergency stop, which calculates L2, L7 and L8 for each agent).

The kit, part by part: what each part contains and what it's for. The templates to fill out are in the Word document and the Excel workbook.

1

Why this kit?

1 record

1.4 Fact Sheet

Why is this the client's business?

The CISO analyzes threats, chooses protections, and monitors, but three decisions remain outside their control. The scope: what data the AI ​​can read, for which users, and whether it can act. Written acceptance of the residual risk, because no measure can eliminate risk entirely. Coordination: according to IBM, only 19% of organizations report coordination between governance and security.

3

Step 2: Evaluate

3 cards

From score to exposure level

The profile score, from 0 to 36, gives four levels: from 0 to 8, low, the eight red lines still to be checked; from 9 to 17, moderate, the CISO and DPO being at least informed; from 18 to 26, high, the likelihood of each relevant risk being raised by one point and a restricted pilot planned before extension; from 27 to 36, critical, with a formal risk analysis by the CISO before any decision.

Four tips for getting a good rating

Rate what can reasonably happen, not the worst imaginable; rate the current situation, without the planned measures, which count in step 3 in the residual risk; justify in one line each deviation from the default value; have the rating reviewed by the CISO, who knows the past incidents and the actual state of the protections.

3.4 Fact Sheet

Read the heat map

The map places each risk in a 4x4 grid, with severity in rows and likelihood in columns, ranging from green to red. It is read starting from the top right corner, where the critical risks to be addressed first are located. The objective of step 3 is to shift the risks down and to the left: the measures primarily reduce likelihood, and sometimes severity.

4

Step 3: Mitigate

14 cards

For each of the fourteen risks, the toolkit proposes measures with a designated person responsible and an effort level (low, medium, or high); those addressing high and critical risks are prioritized. A risk that remains critical after mitigation requires written acceptance from a designated person responsible.

R1

Data leakage via prompts

Provide an enterprise version of the tool, publish a written "what never sticks" rule, implement sensitivity labels and DLP rules, brief all users 30 minutes before access is granted, and review usage quarterly. The rule includes six categories: personal data, identifiers and secrets, source code, unpublished financial data, strategy documents, and information subject to legal secrecy or confidentiality agreements.

R2

Sharing

You don't connect an AI that reads everything based on access rights that no one has reviewed: inventory of sites and folders that are too open, temporary restriction of sensitive sites (Restricted Content Discovery in Microsoft 365), correction of rights by site owners with a deadline, sensitivity labels on HR, finance and management content, pilot on a specific scope before any extension.

R3

Shadow AI around the project

Regulate rather than prohibit, in five steps: identify uses without sanction by an open declaration for 15 days, offer a validated alternative at least as good, publish a catalogue of authorized tools with a response to any request within 10 working days, detect residual uses by the network and the browser (CASB, filtering), then communicate the rule and measure.

R4

Prompt injection

We cannot prevent an attacker from writing an email; we can limit what AI does with it. Treat all content read by AI as untrusted, limit sources to the bare minimum, prohibit automated actions triggered by read content, filter output before it reaches another system, and require the provider to describe its defenses and its remediation timeframe.

MCP Agents and Servers

The client formalizes the agent's mandate: action by action, specifying what the agent can do independently, after human validation, or never. By default, any irreversible action requires validation or is prohibited. Additional features include a dedicated account with task-specific permissions, a whitelist of MCP tools and servers registered with the IT department, a sandbox environment, a log with emergency shutdown, and a review process for each extension.

The 10 rules of generated code

Competent human review before merging, automated security analysis (SAST) and dependency analysis (SCA), no secrets in prompts or code, checking each dependency against slopsquatting, security instruction file provided to the wizard (OpenSSF), AI code tracing, targeted tests on cross-site scripting and log injection, no access to production data, enterprise version, developer training.

R7

AI Supply Chain

Nothing enters the project without being inventoried and verified: inventory of MCP servers, extensions, models and packages; register of approved connectors, the others being blocked; verification of packages before installation; monitoring of component vulnerabilities; models and MCP servers only from sources signed or validated by the IT department.

R8

Supplier and data

What is not written is not guaranteed: the 25-question questionnaire is returned in writing before signing; the contract provides for non-reuse of data, EU location, the retention period of prompts, notification of incident within 48 hours, audit, reversibility and deletion at the end of the contract; the supplier is reviewed annually.

R9

Impact hallucination

AI provides information or recommendations; beyond a certain threshold, a human makes the final decision. Map the decisions where AI intervenes (information, recommendation, decision), require human validation beyond a certain amount or impact threshold, display the sources and level of confidence, test the quality on a sample before deployment and then monthly, and train users to verify.

R10

Fraud and deepfake

No voice, no video, no email alone is enough to trigger a payment or access: double validation off-channel by calling back a known number, oral password or code for urgent requests from the manager, awareness of deepfakes with concrete examples, one-click reporting, semi-annual fraud attempt exercise.

R11

Invisible incident

See, alert, know who to call: log prompts, responses, agent actions and data access for a defined period with the DPO, set up alerts on abnormal volumes, exfiltrations and out-of-scope actions, identify an on-call contact for the tool, prepare quick reference guides and hold a tabletop exercise before the go.

4.3 Fact Sheet

The 10 questions to ask the CISO

Ten questions for a first one-hour meeting, with what a solid answer contains: analysis according to the organization's risk methodology (e.g., EBIOS RM), accessible data and rights, injection protection, logging, leak plan, agent compartmentalization, vendor evaluation, shadow AI detection, pre-production testing, and what the CISO needs from the sponsor.

4.4 Fact Sheet

10 questions to ask the DPO

The legal basis, the need for an impact assessment, the role of the supplier (subcontractor or joint controller), the hosting and storage of data, including prompts, the information of individuals, requests for access or erasure, the retention period of logs, automated decisions, notification to the CNIL within 72 hours and the blocking points, to be known at the scoping stage rather than on the eve of the go.

4.5 Fact Sheet

25 questions to ask the supplier

Questions on data, security, AI security, traceability, incidents, continuity, reversibility, compliance, agents, and governance must be submitted in writing with supporting evidence before signature: an oral answer will not count. A grade does not replace judgment: a "No" to question 1, regarding data reuse for training, carries more weight than several "Partial" answers on governance.

5

Step 4: Decide

1 record

5.7 Fact Sheet

The tabletop exercise

One hour, six participants, and a fictional scenario (leak by a provocateur or unauthorized agent) before the go-ahead: this determines whether the L5 red line is lifted. Five questions to address: how is the leak discovered and how quickly, who decides to cut and how, what data is affected, who to notify and within what timeframe, and what is communicated and to whom. The deliverable is a one-page report outlining three actions.

6

After the go

3 cards

6.1 Fact Sheet

The 90-day review

No later than 90 days after the go, an hour with the sponsor, the CISO, the DPO and a user representative: conditions of the go met or not, reading of the five indicators, incidents and lessons, evolution of the scope, then decision confirmed, adjusted or suspended and date of the next review.

6.2 Fact Sheet

Six indicators to monitor

The percentage of trained users, the number of incidents and reports, the proportion of agent actions validated by a human, the number of shadow tools detected and rectified, the processing time for a tool request (commitment of 10 business days), and the proportion of agents registered with an identity and owner are all tracked. A project that receives no reports in three months is rarely an incident-free project; it's more often a project where no one knows who to report to.

6.3 Fact Sheet

When to repeat the route

Steps 2 to 4 are repeated with each expansion of the scope (especially the transition from reading to action), with each change of supplier or model, after any significant incident or public incident on a comparable tool, with each new regulation (high risk obligations of the AI Act, French law transposing NIS2), and every six months.

Appendices: Ready-to-adapt templates

2 cards

Annex A

Note to the management committee

A page dedicated to three decisions that fall under the committee's purview: the scope, the resources for security measures, and the acceptance of the remaining risks. It presents three IBM 2026 figures, the project evaluation with the toolkit, what the committee must arbitrate, and the cost of inaction: a shadow AI that is progressing, data leaks, and an improvised incident response.

Appendix B

36-point checklist

The thirty-six most important measures of the kit, divided by stage. An item not checked on the eve of the go-ahead must be either a condition dated at the go-ahead, or an accepted residual risk.

Download the AI ​​project sponsor's security and data kit (Word + Excel, free)

65 pages, the four-step process and six appendices, plus the 10-tab tool workbook. Written by NextStart.AI based on publications from ANSSI and CERT-FR, OWASP and CoSAI repositories, IBM and Veracode 2026 reports, and AI projects we support in large organizations.

  • The fourteen security and data risks of an AI project, including those of the agents, explained in client-side language
  • The project profile in twelve questions, the risk rating and heat map
  • 85 measures, questions to the CISO, DPO and supplier, the agent register
  • The eight red lines, the go grid, conditional go, no go, the decision sheet and the incident response sheets

    Associated supportGoverning your AI agents and systems →Registry of AI systems and agents, validation circuit, AI Act qualification and AI committee, set up in six weeks to decide project by project.6 weeksFront door in 2 daysConsulting mission

    8. Frequently Asked Questions

    Is banning consumer AI enough?

    No. According to IBM's 2026 report, the proportion of incidents involving shadow AI has doubled to 43%, with an average cost of $5,39 million compared to $4,63 million for other types of incidents. Shadow AI arises from a legitimate unmet need: the solution is to offer a validated alternative, a catalog of authorized tools, and a fast-track request process, following the principle of "regulate rather than prohibit."

    Is the enterprise version of Microsoft 365 Copilot secure by default?

    It's secure on Microsoft's side, but it inherits your access rights: anything a user can technically open, Copilot can retrieve and summarize. Hence the importance of addressing over-sharing before deployment. The EchoLeak vulnerability of 2025, patched by Microsoft, also demonstrated that the content read by the assistant can be malicious. Copilot Chat, included in Microsoft 365, only queries the web and the documents the user provides, whereas Microsoft 365 Copilot (often called "Copilot Premium") queries the entire enterprise network; and as soon as your teams create agents (Agent Builder, Copilot Studio), the risks associated with agents are compounded.

    Is a no-go a failure?

    No, it's a documented decision. The toolkit offers two alternatives: reduce the scope (less data, fewer users, read-only instead of actionable) or change the solution (enterprise version, EU hosting, different provider), then repeat the process. A "no go" at the scoping stage is much less costly than an incident after deployment.

    Do you need a CISO to use this kit?

    No: it is written for the client. But it does not replace the risk analysis of the CISO; it prepares the discussion with him, thanks to the ten questions to ask him, and the decision sheet includes his opinion alongside that of the DPO.

    How does this differ from the GDPR kit?

    Le sponsor's GDPR kit This kit addresses the compliance of personal data processing: legal basis, impact assessment, and informing individuals. It also covers the security of the project and all its data: leaks, prompt injection, agents, code, vendors, and incidents. The risk of non-compliance (R12) explicitly refers to the GDPR kit, to the AI Act kit and AI and Social Dialogue Kit.

    And what about an agent acting alone?

    This is where the L2, L7, and L8 red lines matter most: no irreversible action without human validation, no agent who fulfills the lethal triad without validation, and each agent registered with a unique identity, owner, and tested emergency stop. The kit adds a level of autonomy defined in the mandate, the lowest privilege level, a whitelist of tools and actions, a sandbox for testing, and a log of every action. An agent acting alone never receives full go on the first pass: they start with conditional go, with a 30-day review.

    Can we allow a desktop agent or an agentic browser?

    Not currently in production, except for specific analysis. These tools control the workstation, files, and browser using the user's session: they almost always combine the lethal triad. On April 13, 2026, the ANSSI's CERT-FR recommended against deploying them on workstations until they are thoroughly tested, and to limit their use to isolated test environments, free of sensitive data, with the approval of the IT department and the CISO.

    Our professions involve vibe coding: should we ban it?

    No, it needs to be regulated, like shadow AI. The risk arises when a tool created in a few hours receives real data and other users without review. The toolkit establishes five rules: declaration, hosting on a validated platform, access control, no real data before review, a designated owner, and an end date.

    Who signs the decision form?

    The sponsor, because the decision rests with them. The CISO and the DPO provide their input, and each critical risk, even after mitigation, must be accepted in writing by a designated manager.

    What should be done if the supplier does not respond to the questionnaire?

    A lack of response is worth zero points in the "8 Supplier" tab. Below 30 out of 60 is a red line to be discussed with the CISO, and without a written commitment not to reuse the data for training, the L3 red line is not lifted.

    9. Going further

    This kit is the "security" counterpart to the other kits from the sponsor: the sponsor's GDPR kit for the compliance of personal data, the AI Act kit from the sponsor for AI literacy, transparency and high risk, the AI and Social Dialogue Kit for the information and consultation of staff representatives, the frugal AI kit for the project's environmental footprint and the responsible AI kit for bias, transparency, and human oversight. To identify and address unauthorized uses of AI, the Shadow AI kit It offers an anonymous survey, an aggregated technical diagnosis, and a 90-day exit plan. AI Systems Governance Kit provides the register of systems and agents on which the 90-day review and reassessments are based. Organizations that want to prepare the decision with their CISO, DPO, and business stakeholders in half a day can rely on the mission Governing your agents and the AI training in companies.

    Read also : GDPR Kit for AI Project Sponsors, Sponsor's AI Act Kit: The 2026-2027 Roadmap, AI toolkit and social dialogue of the AI ​​project sponsor et AI systems governance: the 2026 toolkit.

    The entire series of free kits. To drive adoption: AI charter template, AI training plan template, AI Ambassadors Kit, Manager's toolkit for dealing with AI, AI-powered time-saving kit, Soft Skills Kit for AI, Adoption Roadmap Kit, AI maturity grid, AI Measurement and ROI Kit, AI Governance Kit, Cowork kit, 50 Copilot M365 prompts, Copilot Studio kit, Deploy Microsoft 365 Copilot kit, Copilot Chat Kit, Copilot Excel kit, Copilot Teams and Outlook kit, Copilot Word and PowerPoint kit, Configure Claude kit, Deploying Claude in the Workplace Kit, Claude Agents Kit, Claude Skills kit, Claude's Marketing Skills, Skills Claude HR, Claude Plugins Kit, Claude Code kit, Claude Design kit et Vibe coding kitFor the AI ​​project sponsor: AI Act kit, GDPR kit, Shadow AI kit, AI and Social Dialogue Kit, frugal AI kit, AI Sovereignty Kit et responsible AI kitBy profession: AI HR kit, AI marketing kit, AI finance kit, AI kit for lawyers, AI kit for salespeople, AI toolkit for project managers, AI innovator kit, communicator's AI kit, AI Procurement Kit, AI Kit for Researchers et AI kit Customer service.

    10. Sources

    The figures and positions cited on this page and in the kit are based on the following dated sources. Information current as of October 2026.

    Back to top
    All NextStart.AI offersEight consulting missions, seven pillars, around twenty formats, eighteen professions, five tools
    Around thirty major organizations were supported.Large groups, public agencies and networks: from acculturation to production

    Insurance and social protection

    Public sector and research

    Consumption and retail

    The NextStart.AI collectiveConsultants, trainers and industry experts: more than 5 350 trained professionals since 2023
    All free resourcesDiagnostics, Finders, prompt library, kits by profession, by tool, and for managing an AI program
    The NextStart.AI blogComparisons of tools, practices, agents, governance: analyses and feedback from the field
    Let's talk about your AI projectResponse within 24 business hours
    AI agents in businessUnderstanding, choosing the tool, training the teams, deploying
    Delegate entire tasksClaude Cowork, Copilot Cowork, ChatGPT Work, Vibe Work