Shadow AI: programmable governance to turn creativity into a secure advantage

The massive arrival of code assistants and smart agents to work tools has changed a dynamic that was previously the prerogative of engineering equipment: now anyone can write, deploy and connect code in minutes. This phenomenon - often called "wild code" or "shadow AI" - is not just a technical challenge, it is an organizational challenge that conflicts productivity, visibility and governance.

In recent events, security leaders from companies such as Datadog, Jamf and ASOS have described how the temptation to experiment with models and automations pushes people to draw formal processes. The anecdote of the moderator who burned Claude's tokens all weekend sums up an uncomfortable truth: the IA tools are irresistible, and when that happens without adequate guards, code, infrastructure, and off-review data may contain sensitive information.

The implications are multiple. First, there is a real risk of exfiltration or accidental exposure: scripts that connect services without sanitation, agents with misdefined credentials or public assets without control. Second, traceability and responsibility are diluted: who answers if an automation created by marketing access a critical database? Third, current control frameworks show limitations: OAuth controls or granular permissions often do not express policies by label or context, leaving gaps that agents and users exploit.

In the face of this, the discussion among practitioners converts into a central idea: policies on paper are not enough. Governance must be moved to infrastructure, automate it and make it verifiable. This means implementing data, access and telemetry agents, applying semantic labels to information sets and codifying rules that block dangerous actions without relying exclusively on human surveillance.

From the operational point of view, there are three levers that security teams are prioritizing: first, the classification and rigorous labelling of data as a basis for all control decisions; second, an attractive "governed route" for creators (for example, an internal marketplace of skills or validated templates) that offers productivity without sacrificing visibility; and third, a record of usercases and agents that treat each automation as an asset with owner and purpose. These are practices already recommended by risk management frameworks for IA and applicable security, such as those developed by public institutions and the security community ( NIST AI Risk Management Framework, OWASP).

However, technical implementation must also meet practical limits: cloud and SaaS service providers still offer low-granularity controls for attendees connected to corporate accounts, forcing organizations to complement these controls with their own logical doors, request proxies and data filters that clean or mask information before reaching the model.

Another critical piece is the design of permits: instead of authorizing full integration, organizations need mechanisms to express policies by context - by mail tags, by dataset scope or by tenant - and to force attendees to operate within those limits. This requires additional engineering and sometimes a dialogue with suppliers to request more fine OAuth permits or custom filtering capabilities.

Culture and training cannot be an appendix: the best way to reduce Shadow AI is to make the governed path easier, faster and more visible than the alternative. Investment in safe templates, specific examples by function (HRR, finance, marketing) and timely training reduces the motivation for using personal solutions that are beyond control.

In terms of detection and response, the strategy should include continuous monitoring of artifacts (repositories, deployments, public endpoints), analysis of agent behavior and correlation with human identity in order to track the chain of responsibility. When an incident occurs, a well-maintained user-case record turns a slow forensic investigation into a quick reconstruction of who created what and why.

For CISUS and product managers, the recommendation is clear: prioritizing instrumentation and programmable governance over the ban, creating incentives to use approved channels, requiring identification and owners for each agent, and negotiating with suppliers improvements in permit controls. Implementing these actions now reduces future costs in response to incidents, regulatory exposure and loss of internal confidence.

The "wild code" problem is already within the organization; it is not about turning off creativity, but about channelling it. Companies that make the safe way more useful and simple than the alternative will be those that transform this threat into a competitive advantage, while maintaining the traceability and security required by the connected and regulated world in which we operate.