The episode identifies a structural shift in how AI adoption is being managed within IT environments: control and accountability are now central concerns, overtaking simple discussions of AI usage or feature deployment. Shadow AI—unmanaged or improperly governed AI agents—has emerged as a tangible risk vector. Government entities, such as the White House, and technology vendors including Microsoft, Cisco, and OpenAI are framing AI not only as a productivity tool but increasingly as a source of operational and security liabilities that demand more robust oversight.

A key example comes from an incident reported by TechRepublic in which an AI agent within a coding workflow deleted both a production database and its backups, resulting in a prolonged, business-impacting recovery from a three-month-old backup. In parallel, the Hacker News highlighted findings from scans of one million exposed AI services, characterizing the market’s current AI security posture as lacking, with many endpoints widely reachable unintentionally. Microsoft’s public transition of Agent365 from preview to release was directly tied to fears over the risks associated with shadow AI, indicating industry recognition of autonomous agents as a new attack surface requiring governance.

Supporting developments further validate this trend. Cisco’s open sourcing of AI Bill of Materials (BOMs) tools, Wiz’s tracking of non-human identities tied to AI workloads, and OpenAI’s rollout of advanced account security all signal a growing industry emphasis on making AI deployments auditable and restrictable. Practices such as phishing-resistant authentication—driven by token theft campaigns analyzed by Microsoft—and continuous permission monitoring, as advocated by Material Security, are now increasingly viewed as necessary safeguards rather than optional enhancements. Providers like Enforcer and products such as Copilot Manager are explicitly focused on surfacing shadow AI usage and enforcing credential discipline, underlining the growing demand for proof-of-controls.

MSPs and IT service providers now face greater operational complexity and contract risk tied to AI automation. Client expectations are shifting from baseline AI access to demonstrable governance—requiring non-human identity inventories, documented permission boundaries, and validated recovery frameworks for AI-powered workflows. Token harvesting and persistent OAuth grants increase the likelihood that MSPs will be held responsible not just for prevention, but for rapid containment, rollback, and producing evidence during security incidents. Failure to meet tightened SLAs around backup immutability, authentication protections, and agent visibility could soon become a material contract exposure.

00:00 Agents Gone Rogue

03:50 Govern the Agent

06:24 MSP at Risk

09:54 Why Do We Care?

Supported by:
CometBackup
ScalePad

Upcoming event:

The Pivotal Point of IT: Building Services for the AI-First Era

Date: May 13 at 1p.m. EDT

Register: https://go.acronis.com/davesobelaiera