Tines guidance on governing employee-built AI workflows and agents ('wild code') with built-in security controls
This vendor-authored article from Tines, published on The Hacker News, discusses "wild code": AI-assisted apps, agents and automations that employees build faster than security teams can own, secure and maintain them. It describes no specific incident, vulnerability or threat actor. The risks it names are excessive permissions, sensitive data exposure, unintended actions, orphaned access after role changes, and gaps in investigation. The author argues that education, sandboxes and AI steering committees are insufficient on their own, and recommends building controls into the environment where employees create and run their work. The piece ends with a promotion of Tines and a downloadable guide.
Detalhes técnicos
- Excessive permissions: workflows run with credentials broader than the task needs.
- Data exposure: agents are connected to unneeded data, outputs are shared too widely, and sensitive data can leak into model requests or logs.
- Unintended actions: agents change records or permissions without an approval boundary.
- Orphaned access: workflows keep running after their creator leaves or changes roles.
- Investigation gaps: security cannot readily tell which identity ran a workflow, what it accessed or what changed.
Remediação
- Govern before deployment and keep governing afterward: define permitted data sources, actions and access up front, scope permissions to what the workflow needs, and require human approval for consequential actions. A prompt instructing an agent is not an access control.
- Make the approved route the easiest route, with accessible building tools and approved connections, so employees do not turn to workarounds.
- Design for investigation and intervention: keep ownership, execution identities, actions and changes traceable, and avoid capturing sensitive data in logs.
- Establish in advance who can pause a workflow and how its access is revoked.
- Isolate workflow steps and inject credentials at runtime rather than exposing them in code.
Linha do tempo
-
Tines guidance on governing employee-built AI workflows and agents ('wild code') with built-in security controls
ler a análise desta fonte ocultar
This vendor-authored article from Tines, published on The Hacker News, discusses "wild code": AI-assisted apps, agents and automations that employees build faster than security teams can own, secure and maintain them. It describes no specific incident, vulnerability or threat actor. The risks it names are excessive permissions, sensitive data exposure, unintended actions, orphaned access after role changes, and gaps in investigation. The author argues that education, sandboxes and AI steering committees are insufficient on their own, and recommends building controls into the environment where employees create and run their work. The piece ends with a promotion of Tines and a downloadable guide.
Technical Details
- Excessive permissions: workflows run with credentials broader than the task needs.
- Data exposure: agents are connected to unneeded data, outputs are shared too widely, and sensitive data can leak into model requests or logs.
- Unintended actions: agents change records or permissions without an approval boundary.
- Orphaned access: workflows keep running after their creator leaves or changes roles.
- Investigation gaps: security cannot readily tell which identity ran a workflow, what it accessed or what changed.
Remediation
- Govern before deployment and keep governing afterward: define permitted data sources, actions and access up front, scope permissions to what the workflow needs, and require human approval for consequential actions. A prompt instructing an agent is not an access control.
- Make the approved route the easiest route, with accessible building tools and approved connections, so employees do not turn to workarounds.
- Design for investigation and intervention: keep ownership, execution identities, actions and changes traceable, and avoid capturing sensitive data in logs.
- Establish in advance who can pause a workflow and how its access is revoked.
- Isolate workflow steps and inject credentials at runtime rather than exposing them in code.