Use Cognizant’s open sandbox to cut risk of rogue agents

The HFS Highlight is for CIOs, chief AI officers, and enterprise operations leaders testing how multi-agent systems behave before production deployment.

Handling unexpected emergent behaviors in multi-agent systems is a problem HFS first identified in our work with Cognizant for the paper Design your agent OS to win the AI future, published in March this year. When given autonomy, AI agents regularly go off-script, creating risks to systems, people-facing processes, and enterprise decision making.

More than one-fifth of Global 2000 enterprises building multi-agent systems identified emergent and unpredictable agent behaviors among their unforeseen challenges when deploying multi-agent systems (see Exhibit 1).

Cognizant responded with a pre-deployment risk tool, TerraLingua, and has now made it available to the public. It’s a timely and relevant move, just as the market is spooked by news that an OpenAI model became so fixated on solving a narrow benchmark that it found a way to escape OpenAI’s isolated test environment, reach the open internet, and steal credentials to go deep into Hugging Face’s production infrastructure.

Exhibit 1: Unpredictable emergent behavior in agents is among the leading unforeseen challenges leaders encounter when rolling out multi-agent systems

Question: Please share any unforeseen challenges or emergent behaviors you encountered during your agentic solution deployment.
Sample: Leaders from 108 Global 2000 enterprises, optional open-ended question
Source: HFS Research, 2026

Observe agentic behavior in a safe place to learn how to reduce the risks

TerraLingua is a container built to surface unplanned, goal-obsessed agent behavior without real-world consequences. The OpenAI incident is what happens when containment itself fails. By OpenAI’s own account, it was caused by the human error of a configuration mistake, not a rogue intent by AI. The human error allowed a goal-seeking model to reach real infrastructure. But the underlying behavior is the same one Cognizant has spent months cataloging: Agents given a narrow objective and enough autonomy will find paths nobody designed for or expected.

Understanding the risks enables the development of guardrails and controls to reduce them.

TerraLingua is a persistent, multi-agent research sandbox that has been under test for months, led by Babak Hodjat and his team at the Cognizant AI Lab in San Francisco. HFS has seen demos and written about its emergence in the HFS Highlight, Agent-only companies show what’s next for Ops leaders. Anyone with an API key can now build an agent, drop it into a shared environment, and watch what it does next in Observer Mode.

See what happens when your agents interact with agents from other systems

Cognizant has already observed agents foraging for energy, talking to nearby agents, reproducing, and leaving behind notes, warnings, and manifestos that outlive them and shape what later agents find. There is no assigned task, and there is no instruction to cooperate or compete. Left to their own devices, agents have been observed forming rival ideological factions, building durable economies, and, in one alarming example, spiraling a simple food-safety warning into a population-wide panic that starved a colony sitting on abundant resources.

The sandbox provides enterprises building multi-agent systems with a way to know how their agents will behave once they’re autonomous and interacting with systems that nobody had been able to explicitly test for ahead of production.

The Bottom Line: De-risk multi-agent deployments by testing how your agents play with others before you release them in the wild.

The challenges and risks of multi-agent systems only rise as agents interact with agents from other systems, as they will increasingly need to do as firms orchestrate agents from multiple sources to deliver end-to-end.

You don’t get to choose whether your agents surprise you, only whether you find out in a sandbox or in production. TerraLingua won’t stop a determined model from finding a zero-day vulnerability. But it does offer a public instrument for watching the failure play out behind closed doors before it does the same in public with your logo attached.