Brief IA
›

AI Providers: Assess Usage Before Granting Access

Business & Startups·Tom Levy·

AI Providers: Assess Usage Before Granting Access

AI Providers: Assess Usage Before Granting Access
Key Takeaways
1The same provider can present different risks depending on the deployed capabilities and their usage.
2The proposed checks cover data, agent permissions, reliability, dependencies, and contracts.
3A change in model, permissions, or purpose should trigger a reassessment.
Why it matters : the review focuses on accessible resources, possible actions, and the consequences of failure, beyond the general security of the provider.

For information system security managers, the evaluation of an AI vendor should focus on the intended deployment, not just its overall security. The recommended approach combines tailored checks based on exposure, conditional authorization, and post-deployment monitoring. A conventional questionnaire is insufficient to cover these dimensions.

Management Must Assume Accepted Risk

Authorization should depend on sufficient justifications, operational controls, and a residual risk compatible with the company's tolerance. It could also be conditional: categories of data, users, connections, agent rights, or permitted uses would then be limited and documented. The purpose of the decision would be the acceptability of a deployment aimed at a specific commercial objective.

Missing essential elements or unfinished corrections would justify a delay, with a designated responsible party and a deadline for re-evaluation. A refusal would be necessary when the risk cannot be sufficiently reduced, when the vendor fails to meet essential obligations, or when the remaining exposure exceeds the accepted threshold. For each use case, the decision should retain its justification, conditions, a review date, and criteria triggering an exit.

The acceptance of residual risk and significant exceptions should fall under the purview of executives or authorized delegates, within the limits of their mandate. Security and privacy teams would be responsible for access controls; legal and compliance would interpret obligations. Procurement would handle commercial conditions, while business leaders would need to justify the use and its operational protections.

The Inventory Must Go Beyond Official Purchases

Software marketed without an AI label may still incorporate such functions. Therefore, the scope to be examined should encompass operational tools, customer-facing applications, security products, and online software. It should also cover autonomous generative assistants, on-demand AI platforms, external models, programming interfaces, agents, and development tools.

Vendors, including those providing managed services, would be relevant when they use AI to process company data or deliver their services. The inventory should also extend beyond official acquisitions to include employee experiments. Unauthorized use can expose data, identities, or internal systems.

For each deployment, the review sheet should specify its commercial purpose, users, scope, data categories, and granted permissions. It would distinguish between a result that informs an important action and a system that executes it, while identifying integrations, model providers, and other significant dependencies. A proposed practical approach is to inventory AI among existing vendors, then choose a high-impact use case for which the risk level and authorization conditions will be documented. The insights gained would then allow for the adaptation of the process to both new tools and those already in use.

Controls Must Follow Actual Exposure

Accessible resources, possible actions, and the consequences of a failure determine the risk of a deployment. A tool summarizing ordinary meeting notes would thus require less thorough examination than an agent consulting production or possessing privileged permissions. An initial filtering of exposure should precede requests for additional justifications and set an initial risk level.

Classification should consider the sensitivity of information, access rights, autonomy, potential damages, criticality to the business, regulatory constraints, and the possibility of rollback. A vendor could thus belong to different categories depending on the selected product and its use. Rather than a separate circuit or a uniform requirement, the approach would complement procurement and third-party risk management with AI-specific criteria, proportionate to internal rules and product capabilities. It would cover the vendor, deployed functions, shared information, influenced decisions, and subsequent controls.

La lettre du soirGratuit · 19 h

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

At the proposed Level 1, information that is not sensitive, no privileged permissions, and low-impact results would call for basic checks of security, privacy, and data usage. Level 2 would concern, in particular, confidential business information, significant dependencies, or results requiring reliability and human oversight; the review would also focus on model governance, recorded traces, contractual guarantees, and recovery options. Level 3 would target regulated sensitive information, privileged agents, significant effects on customers, or decisions with substantial financial, legal, security, or operational consequences. Specialized expertise, documented mitigations, and acceptance by management would then be required.

Evidence Must Go Beyond Commercial Assurances

Standard checks already cover cybersecurity, access, certifications, incident management, continuity, and information protection. They should be supplemented by examining the retention of queries, uploaded files, results, and customer data, as well as their potential reuse for training models. The location, deletion, ownership, and processing methods of these elements would also be among the points to verify. Vendor statements alone do not demonstrate the suitability of their AI for each use: verified elements should be distinguished from their assurances.

Technical analysis should examine identities, client isolation, encryption, agent permissions, accessible tools, and defenses against query injection. Given that results can be false, manipulated, biased, or insufficiently reliable, tests should explore relevant failures and system limits. They would also address response validation, human oversight, and protections against dangerous actions. For the riskiest uses, a pilot, configuration review, or targeted tests should validate critical claims about actual integration and its permissions, rather than relying on a simple demonstration.

The expected documentation should identify foundational models, subcontractors, datasets, and external services, as well as known versions and limits. Audit trails, incident notification procedures, and available independent elements would need to be examined. Contracts should specify continuity, backup solutions, cooperation during an incident, audit rights, portability, and exit conditions; the company should be able to restrict access, investigate, and exit the service. The results of checks, responsible parties, correction timelines, and remaining risks would be documented.

Authorization Must Remain Reviewable After Deployment

A model's evolution can change its behavior, performance, or exposure to security risks after its authorization. A change in version, data practices, permissions, or purpose should therefore trigger a new review. Monitoring should also track incidents, failed controls, vendor changes, and emerging dependencies. A workable exit or rollback mechanism should be defined.

Frameworks exist to organize these checks. The AI RMF from NIST addresses the identification, assessment, and management of AI risks; its profile dedicated to generative AI adds considerations regarding vendors and the lifecycle. NIST publication 800-161 Rev. 1 examines the extension of cybersecurity supply chain risk management practices to dependencies and vendor resilience. ISO/IEC 42001 provides guidance for evaluating AI governance within organizations.

These frameworks complement the review of use, contracts, and applicable legal or regulatory advice, without replacing them. Existing platforms for governance, risk, compliance, and third-party management could be used to gather evidence, process authorizations, track corrections, and organize re-evaluations. AI-specific questionnaires and checks could be added.

Brief IA — L'actualité IA en français

L'essentiel de l'actualité de l'intelligence artificielle, décrypté et expliqué chaque jour.