Shadow AI to Shadow Agents: What's Actually Changed in 2026?

Key frontier models are now capable of spotting vulnerabilities that no one thought even existed. They are finding software vulnerabilities at scale without any prompting, presenting both opportunities and challenges for security teams. For instance, recently, the Associated Press reported on the Anthropic Mythos model, spotting vulnerabilities in highly sensitive U.S. government computer systems during a testing exercise. Security researchers describe this ability as dual-use.

The Mythos model, part of Anthropic’s project Glasswing, partnered with national intelligence agencies to find and fix vulnerabilities in critical systems before attackers get to them, as per AP. The frontier AI model “broke into almost all of our classified systems, not in weeks, but in hours,” Joshua Rudd, National Security Agency (NSA) chief, seems to have informed Senator Mark Warner of Virginia.

This kind of scanning capability helps defenders patch vulnerabilities faster, and also provides attackers with a more effective tool. As our guest noted in the recent The Security Strategist podcast episode, it's "almost like a new weapon." Both sides of the security battle believe they can leverage it to their advantage.

This tension sets the stage for a conversation that has been growing in enterprise security circles throughout the year – Shadow AI.

In this episode of The Security Strategist podcast, host Shubhangi Dua, Podcast Producer and B2B Tech Journalist at EM360Tech, sat down with Guru Sethupathy, Head of AI Governance at Optro, to break down Shadow AI and its looming threats in 2026 and beyond.

Over the last couple of years, Shadow AI has mainly involved employees entering sensitive data into ChatGPT while IT teams rushed to respond. This summer, the narrative has changed. The EU AI Act's regulatory clock is ticking. Meanwhile, Shadow AI is transforming into autonomous "shadow agents" that operate without waiting for human approval at every step.

Sethupathy believes many enterprises are about to realise they can't comply with regulations for systems they don’t even know they're using.

Rogue Chatbots to Rogue Agents

Shadow AI isn’t a new concept, but Sethupathy challenged the notion that enterprises have been aware of it for years. "Even into 2024, Shadow AI was not the main priority for enterprises," he said. It only gained traction heading into 2025, and "it has dramatically increased in 2026."

Two factors are driving this shift, he explained. The first is the "democratisation of AI." Any employee within an enterprise can now access and interact with AI tools directly. The second factor is the countless ways AI can enter a company. AI can come through third-party vendor tools, web browsers, internal development, or no-code and low-code platforms. When you combine these entry points with the number of employees interacting with them, he noted, "you can see why there’s so much AI in an organisation that isn’t being tracked."

However, the real change, Sethupathy argued, isn’t from chatbots but from what followed. "With chatbots, companies could exert control," he said, highlighting the relative ease of restricting the tools staff could use and controlling how data flowed through them. AI Agents present a different challenge entirely: "It’s not just about data leakage. It’s not just about data security. These agents are taking actions, making decisions, and acting."

This, he said, poses "a level of risk that is much higher."

Why is Governance Failing Against AI?

If agentic AI is the new area of risk, why hasn’t governance caught up? Sethupathy attributes this to tempo. "Governance in the past has typically been a point-in-time exercise," he said. But autonomous agents operate constantly. They learn consistently and take actions, access tools, and data around the clock.

"Imagine you’re driving, and your car only informs you of your speed every half hour. That would be pointless,” the Head of AI explained. The same reasoning applies to agents, he argued. With their continuous work, oversight needs to be continuous as well. This creates two major problems for organisations: a process issue (rebuilding governance frameworks around ongoing review) and a technology issue. A