AI Agents Behaving Badly: Four Security Rules For The Agentic AI Age

Srinivas Shekar, Founder and CEO, Pantherun Technologies.

​Bots now generate over 57% of all web traffic, compared with less than 43% of human activity. This is only set to expand as agentic AI proves its mettle in online shopping, research, cloud administration, customer service and more. But enterprise security systems were designed to give approved human users access while blocking human and machine intruders.

It isn’t just about overcoming the “prove you’re an approved human” paradigm that legacy security systems were built on. It’s also about dynamism. After all, humans have stable identities and permissions that can be established and recognized. But AI agents create subagents and change their behavior according to the task they’re executing. They can communicate and interact like humans, but they can also make API calls, move data, consume machine-readable streams, interact directly with enterprise infrastructure, create software tools on the spot and perform thousands of operations simultaneously.

In a world where malicious bots and benevolent agents co-exist, how do we tell them apart? And what happens when we can’t?

AI Agents Behaving Badly

Malicious AI agents are already getting into places they shouldn’t be and executing actions. In July 2026, the agentic ransomware attack known as JadePuffer exploited a flaw in the AI development tool Langflow to conduct reconnaissance, steal credentials, encrypt files and demand a bitcoin ransom for release. Once inside, the agent determined what opportunities to exploit, attacked several avenues simultaneously, generated attack code dynamically and adapted its own code in real time, until it achieved success. The agent also created software on the fly, allowing it to move ransomware from a compromised container onto the host. It even wrote its own ransom note!​