CryptoSlate
AI agents used each other's arithmetic results from a shared workspace without being told that other agents were present, a Nov. 9 preprint found. With an instruction to recheck the workspace, GPT-6 Astra pairs succeeded in 49 of 50 and 47 of 50 trials across two shared-store implementations. Both private-store controls produced zero successes in 50 trials.
Why it matters
Pekka Abrahamsson and seven Tampere University co-authors tested agents on tasks in which two held complementary halves of an arithmetic problem. Each could calculate a partial result and use the store to retrieve the missing value. The prompts described storage as a convenience for intermediate work, not as a shared space, while task-supplied variable names gave agents clues about what to find.
The instruction to check again before submitting made a substantial difference. Cross-agent access occurred in all 50 shared-store trials in each implementation with that instruction, versus one in 50 without it. The experiment demonstrates access and useful influence, not an intent to communicate.
Market impact
For businesses deploying AI agents, the finding highlights shared caches, scratch directories and memory services as potential channels between otherwise separate tasks. The authors recommend separating storage when sharing is unintended, using namespaces where appropriate and logging authorship.
The arithmetic setting was small, and results varied by model and implementation. The study does not establish how often unintended sharing occurs in deployments.
Frequently asked questions
-
How did the researchers test shared access?
Two agents received complementary halves of an arithmetic problem and could use a workspace store for intermediate results. A control kept the interface unchanged but made the store private to each agent.
-
What difference did private storage make?
With private stores, both implementations produced zero joint successes in 50 trials. The missing information could no longer arrive from another agent through the store.
-
How did the workspace-recheck instruction affect access?
Cross-agent access occurred in all 50 shared-store trials in each implementation with the instruction, compared with one in 50 in each without it.
-
Did the experiment show that agents intended to communicate?
No. The experiment demonstrated access to another agent's information and useful influence on task results, but did not establish an intent to communicate.
-
What safeguards did the researchers recommend?
The authors recommend separating storage where sharing is unintended, using namespaces where appropriate and logging authorship.
TheBlock
Crypto News
WuBlockchain
CoinDesk
CoinTelegraph
WatcherGuru
Altcoin Daily