Study: A Few Rogue AI Agents Can Steer Entire AI Crowds

Many of us now use AI assistants that don't just answer questions — they take actions. They book flights, compare prices, reply to emails, and increasingly talk to other companies' AI assistants. Researchers at several universities asked a simple but uncomfortable question: if some of those assistants are controlled by bad actors, how many are needed before the whole group starts behaving badly?

The standard answer has always been 'a big chunk of them.' This paper says that's wrong. Instead of attacking head-on, a small group can nudge the crowd through middle steps — intermediate positions that are easier to reach — and end up somewhere the majority would never have agreed to directly. Think of a lobbyist who never asks for the extreme policy, just the halfway one, then the next halfway one. Each hop is cheap; the destination is the same. In the researchers' experiments with AI agents, these 'stepping stones' cut the number of bad actors needed and let them skip majority approval entirely.

Timing matters too. Attack when a group is uncertain or distracted, and the same crowd can be pushed farther. So an AI population's resistance to manipulation isn't a fixed property — it depends on the whole landscape of other behaviours the group could settle into, and how easy those are to hop between. That's genuinely new thinking, and it applies to things you might actually use: a swarm of shopping agents that could be steered toward worse deals, or news-summarising assistants nudged toward a particular story.

The honest catch: this is a research finding, run in simulations and maths, not a documented attack on real products. Nobody has hijacked your assistant yet. But if the conclusion holds, checking each AI model individually isn't enough. Safety will also mean mapping how these agents influence each other — the digital equivalent of watching the crowd, not just the person.