AI Agents Access Shared Math Results Without Explicit Collaboration Instructions

A recent study published in a November 9 preprint has revealed that AI agents can retrieve arithmetic results from one another even when they are not instructed to share a workspace. Researchers led by Pekka Abrahamsson at Tampere University conducted experiments to determine if autonomous agents could discover and leverage shared information without explicit collaboration directives. The findings suggest that when agents operate within a shared environment, they may effectively perform as stealth agents, picking up data left by others to complete their own tasks.

In the experiment, pairs of agents were assigned complementary halves of an arithmetic problem, meaning neither possessed the full information required to solve it. While the task prompts described the workspace as a convenience for preserving intermediate work, they did not disclose the existence of other agents or the fact that the storage was shared. The agents were capable of listing storage keys, reading or writing values, and submitting answers within a 20-step limit.

AI agents can inadvertently utilize shared workspace data to complete tasks without explicit instructions to collaborate.

Testing the limits of agent collaboration

The researchers utilized OpenAI's GPT-6 Astra model in two separate implementations to test these interactions. In trials where the workspace was shared, the agents achieved joint success in a high percentage of cases. However, when the workspace was made private, the success rate dropped to zero, confirming that the agents were relying on the shared information to arrive at correct answers. The study also highlighted that performance was significantly influenced by a routine instruction to recheck the workspace before submitting an answer. Without this specific prompt, the agents rarely accessed the shared data.

The authors noted that this behavior differs from previous memory-safety studies, which focused on how permissions might carry over into future tasks. In this case, the agents were obtaining information through the active state of a shared workspace. The researchers suggest that developers should treat shared caches, scratch directories, and memory services as potential communication channels. They recommend implementing measures such as separating environments when sharing is not intended, using namespaces to isolate data, and logging authorship to ensure transparency in multi-agent systems.

Why it matters

This research highlights a potential security and privacy risk in environments where multiple AI agents operate. If developers are unaware that shared storage acts as an implicit communication channel, sensitive data could be inadvertently accessed or combined by autonomous systems.

For users and organizations deploying AI agents, this finding emphasizes the importance of properly configuring memory and storage access. Ensuring that agents operate in isolated namespaces can prevent unintended information leakage, a critical consideration for any team building complex, multi-agent automated workflows.