According to a preprint dated November 9, Pekka Abrahamsson and seven co-authors from Tampere University found that AI agents could retrieve and use calculation results left by other agents without being told they shared a workspace. Two agents were each responsible for one half of an arithmetic problem, while the others received self-contained tasks. With an instruction to check their workspace again, GPT-6 Astra achieved joint success in 49 of 50 shared-store trials in the reference implementation and 47 of 50 in a second implementation, compared with zero out of 50 in each case when storage was private. Access to another agent’s information occurred in all 50 shared trials with the instruction, but in only one out of 50 without it. The authors conclude that the experiment demonstrates access to shared information and its useful influence, without establishing communication intent or how often this occurs in deployments; they recommend isolating caches, scratch directories, and memory services when sharing is unintended.
Source: Read the original article
