Security teams must tame wild code as AI workflows outpace governance
Employees are building AI agents faster than security can track them, creating risks around permissions and data exposure that require new controls.
Key points
- AI workflows often run with excessive permissions or expose sensitive data to unintended audiences.
- Orphaned access occurs when creators leave but their automated workflows continue operating without oversight.
- Traditional education and sandboxes are insufficient; controls must be built into the environment itself.
Security teams face growing challenges as employees build AI-assisted workflows that spread faster than governance processes can support them. According to a report by The Hacker News, these tools often become operational dependencies before security teams are aware of their existence or scope.
A sales manager might create an AI workflow to route leads, summarize requests, and post updates to shared channels. While this saves time and reduces manual work, it introduces critical security questions about credentials, data access, and output visibility that traditional oversight methods fail to address.
The Mechanics of Wild Code
The term wild code refers to useful builds that expand faster than the processes for owning and securing them. It is not merely unapproved software; employees are creating new capabilities rather than just adopting existing products. A small experiment can quickly become a critical business function, leaving security teams unable to trace actions or stop further access if the tool misbehaves.
The risks extend beyond vulnerabilities in generated code. Even properly functioning code can operate with too much access or expose data to the wrong audience. Security teams must consider which records the workflow can access and whether its output is visible to unauthorized users.
Operational Risks and Gaps
Excessive permissions are a primary concern, where a workflow needs only a small set of records but runs with credentials capable of reading or modifying far more. Sensitive data exposure occurs when employees connect agents to unnecessary information or make outputs broadly available, potentially leaking data into model requests or logs.
Unintended actions can happen when agents change records or permissions without appropriate approval boundaries. Orphaned access arises when a workflow continues running after its creator changes roles or leaves the organization, leaving no clear owner to review permissions. Investigation gaps further complicate matters, as security cannot readily establish which identity ran the workflow or what it changed.
What should you do?
Employee education, sandboxes, and AI steering committees are helpful starting points but are not enough to control wild code on their own. The mistake is expecting these measures to govern software they may never encounter or to control behavior after the software starts running. A policy alone describes what should happen but does not enforce it in real time.
A more sustainable approach builds controls directly into the environment. This includes defining what employees can connect to, what their apps and agents can do, and how that work is monitored once it is running. Security must shift from passive oversight to active environmental controls.
What to do and how to stay safe: AI security
- Implement technical controls that restrict what AI agents and workflows can connect to within your environment.
- Establish monitoring mechanisms to track what automated tools access and change in real time.
- Define clear approval boundaries for actions that modify records or permissions to prevent unintended changes.
- Create processes to review and update permissions when employees change roles or leave the organization.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
What is wild code in the context of AI security?
Wild code refers to employee-built AI apps and automations that spread faster than security can own, secure, or maintain them.
Why are traditional security measures insufficient for wild code?
Education and sandboxes cannot govern software security may never encounter or control behavior after the software is running.
What are the main risks of AI workflows?
Risks include excessive permissions, sensitive data exposure, unintended actions, orphaned access, and investigation gaps.
