Microsoft: How to Secure Enterprises from Agentic AI Risks

AI powered automation with human-agent teams are redefining the frontiers of work in major companies. 

With vibe coding blowing up, application and agent development has been democratised, opening doors for non-developers with area specific skills to use these low-code tools to build their own agents, thereby rapidly increasing adoption. 

Microsoft’s AI security report titled Cyber Pulse shows that over 80% of the Fortune 500 companies have active agents built using these low‑code/no‑code tools. 

This is contributing to a drastic change in workflows as organisations encourage embedding task-specific AI agents into their everyday workflows. 

“A key difference we are seeing in AI agent adoption from other technology eras (such as early cloud adoption) is the pace and magnitude of change,” Charlie Bell, former Executive Vice President at Microsoft Security writes on LinkedIn.

“AI agents are moving from experimentation to operating at scale at a furious rate. This shift is fundamentally changing the security landscape.”

Agentic AI adoption by industry 

According to the Microsoft report, software and technology industries have the highest rate of agentic AI adoption rated at 16%. 

Displaying a systematic change in workflows across factories, supply chains and energy operations, the manufacturing sector boasts 13% of global agent use.

Financial services sector – which includes banking, capital markets and insurance – are also leading the charge in agentic AI adoption along with the operationally intensive manufacturing sector as well as customer-facing retail sectors.  

As Microsoft’s first-party telemetry from November 2025 reveals, the financial sector accounts for 11% of all globally active AI agents. 

This dominance of the financial sector in AI adoption also has considerable upside, as proven by Barclays wherein it surpassed 1 million hours of productivity using Microsoft Copilot. 

Microsoft’s first-party telemetry puts retail at 9% as AI agents are used to “optimise inventory, enhance customer experience and streamline frontline operations”.

What are Double Agents?  

Research from Microsoft’s AI Red Team discerned how AI agents can be turned against the organisation using many subtle techniques such as embedding harmful instructions in everyday content and redirecting agent reasoning by manipulating task framing.

Thereby arises the conundrum of the double agent – a misaligned AI agent with too much access or wrong instructions which can be used by bad actors to exploit enterprises. 

The risk arising from them is manyfold, including prompt injections and incorrect data permissioning which can cause data leakage.

Sometimes, even a single misdirected file or overshared document could have extreme financial or regulatory repercussions making agent security non-negotiable as adoption amplifies. 

Security concerns soar with Microsoft's Data Security Index revealing that only less than half (47%) of the organisations have implemented Gen AI security controls and over a quarter (29%) of employees use unsanctioned AI agents for work. 

How to securely use agents to “innovate at machine speed”

To secure AI agents, organisations must design systems assuming attackers will get inside.   

The Microsoft report suggests a holistic AI risk management effort that takes a top-down approach with a dedicated executive who is accountable for governance. 

By adopting zero trust principles for AI agents, organisations can verify the identity, device health and location to evaluate the risk level before agents are allowed access. 

Ensuring that agents only receive the privileges needed for carrying out a task can root out privilege escalations.  

This can be done by defining the scope and least privilege required for each agent instead of assigning broad privileges. 

Organisations should apply regulatory data compliance laws to AI channels and develop business continuity playbooks for various AI scenarios. 

“As AI becomes embedded in every workflow, the real differentiator isn’t adopting models – it’s governing them quickly and correctly," the report states.

“Strong identity, data security and oversight for agents and co-pilots are now essential to earning trust and avoiding escalating operational, legal, and reputational risk.

“Leaders who act now to put guardrails in place won’t just mitigate threats – they’ll unlock faster innovation, protect customer trust and build a durable advantage in an AI‑driven economy.”