A communications manager sending an outside writer an ART that he has not published to the public. Borrower accounts for the Poppy Dashboard and pastes the brief into their account to generate a draft. Absolutely no bad stuff goes on and the final product is perfectly readable. But a confidential document has crossed an agency and security red line.
This is a typical form of shadow AI in content operations: valuable work done via services the organization isn't aware of, or is not able to manage or monitor. Includes legal, investor and recruitment relationships and executive teams handling sensitive documents. As mentioned in CISO Platform's article on using AI as a security platform, policy should consider what they use AI for, rather than just what apps they installed on their IT.
Where Shadow AI Enters the Content Supply Chain
Content is seldom associated with its source. Briefs find their way to freelance people, editors, design agencies, translation vendors, SEO consultants, and publishers. Every way it's passed on can result in a duplicated copy.
There's never anything that's not a good fit for unrestricted models, usually something that is just bad. Projects always have info that is not good for unrestricted models, like customer interviews, launch dates, executive correspondence, unreleased pricing, internal research, etc. If an editor or browser extension has global access, a simple request made by users for "making this paragraph clearer" can result in the entire surrounding document getting sent to them.
Take, for example, a business writing an executive book or industry book that will be sent out to buyers. One thing that an author can use is a platform like AIWriteBook, which provides the manuscript import and AI support with regards to writing and editing. This is not an inconsequential choice, because an uploaded draft will have to be treated differently, if it includes confidential customer story information, than a manuscript composed 100% from public databases.
If you're looking for a control, identify the Unapproved Workflow first!
Security teams typically start off by providing a list of what applications they can't stand. Meanwhile, content teams work on the basis of tasks: "research, " "draft, " "rewrite, " "transcribe, " "translate, " "artwork, " "distribute. Often the answer to “Do you use AI?” is a sideshow to an audit of the employee's tasks that has already been performed while mapping those tasks.
Follow One Asset From Brief to Publication
Choose a current campaign and follow its working files. Who is the source of the short text? Who were the external contributors that received it? Was it transcribed into a chatbot, a transcription service or an editing extension? Who had access to drafts and where were these kept? Review personal or Agency-managed systems.
Discovery should encompass both sanctioned and unsanctioned usage via identity, endpoint, browser, network and SaaS telemetry. These referrals are far from perfect, particularly for contractors, and should be supported by interviews and review of contracts.
Rethink Product Discovery as the beginning of assessment
Tools are readily available. For example, Ai Product Index categorises products based on different categories, different features and pricing. A directory does help to compare options, but listing the things does not prove that good data handling. Please check vendor's current terms, model providers, training controls, retention times, subprocessors and deletion process before approval.
Ask students to decide on the types of material, and also on the uses of the material that are allowed.
A practical policy has to address two questions: What information can be fed into a service? What can a service do? There's not much either the blanket "no AI" or "no No-longer-AI" works for.
Low-risk tasks can get a less complex way.
Using generic headlines that are taken from public product descriptions, summarising published articles or checking grammar in those that have been approved can be a lighter approval route. Ensure there are specific accounts for employees to use, and do not enter credentials, unpublished data or any personal information. Work, even that supplied free of charge must be checked for accuracy and ownership.
Restricted Inputs: Enforceable Boundaries
Ban confidential customer data, unreleased financials, authentication secrets, employee data and privileged legal documents from unapproved services. If the deployment is to be done as part of an approved enterprise, set tenant configuration, retention policies, logging and any applicable contractual limitations on training the model, access controls. While there are borderline cases of detections for sensitive content, there are policy exceptions and false positives that must be owned.
The NIST Generative AI profile helps formulate governance and risk-management assignments. It should give a workflow-specific control design, not an excuse for vendor diligence.
Outline how AI tools are addressed in agency contracts.How AI Tools are Covered in Agency Contracts
Outsourcing, however, should not result in relocation of responsibility for the data in its own right. The terms of a marketing agreement typically address when the final drafts are due, whether such drafts are to be revised, who owns the final product, etc., but may not address who helps draft the final product, who retains the rights to any models or accounts used to create it, etc.
That's the point with specialist services including SaaS out there marketing and link developing mainly Get Pro Links. That is not to suggest that any specific provider is employing unapproved AI – but it puts them where everyone else in the content/distribution chain is.
Contract provisions should provide clarity on acceptable uses of AI, the prohibition of unauthorized uploads of sensitive information, the disclosure obligations when an AI system uses a third-party provider as a subprocessor, retention and deletion of data, and responsibilities for notifications. Request evidence that is commensurate to content presented. A vendor who is editing an acquisition announcement intended for public distribution should not be subjected to the same review as a public acquisition announcement made by a vendor editing a confidential announcement.
Look beyond Prompts for Review Outputs and Integrations.
The main reason why this input leakage is given attention is that it is “visible”. It is simpler to overlook the risk of production. An article can be created to present a quote from the customer, incorrect wording of a regulation, questionable claims, or information that is implied in an internal brief. Need external approval – check factual accuracy, checking for quotes, permission, confidentiality, brand commitments.
Technical integrations are treated on a par with others. Defending against an AI writing extension that has access to all the exposed pages can turn out to be a different type of expose as opposed to an application that is processing pasted content. Limiting OAuth scopes, drives permissions, browser permissions, API keys, and automated publishing rights to the task.
A 30-Day Shadow AI Governance Plan for CISOs.
During those first 10 days he gathered input from several interview partners and reviewed two or three projects that were finished, and eventually determined which limited information might lead to a vulnerable company-controlled system. Release an interim public policy guideline that differentiates public information from confidential and/or controlled information. Provide teams with a contact for when they need to escalate instead of ambiguous bans.
Over the coming 10 days, evaluate most-available services and top-session vendors. Provide agency onboarding and statements of work templates to update and provide a safe space for low-risk work, as well as limit risky browser permissions. Take the focus off of controls that address the reoccurrence of unsafe workarounds.
Test with one campaign during the last 10 days. Ensure that contractors are familiar with the rules; maintain records of any exceptions; and help reviewers follow the source files. Report on a few of the operational actions taken: if the organization has adopted approved tools to cover non-standard activities, it would be appropriate to report on this; exceptional high-risk conditions would be reported if they exist; the number of vendor reviews completed; and the time required to approve a legitimate use case.
Comments