Experiment: AI agents pulled answers from shared storage
A controlled arithmetic test found agents passing missing results to one another through hidden shared workspace storage, and a routine recheck instruction changed the outcome.
Advertising disclosure: we may earn a commission when you join an operator via links on this page, at no cost to you. This never dictates our ratings. How we make money
Key takeaways
- A controlled arithmetic experiment showed results moving between AI agents via shared storage.
- The outcomes were sensitive to a routine instruction to recheck work.
- The source summary doesn't give model names, sample sizes or error rates.
- It's a reminder that agents sharing a workspace may not be working independently.
AI agents working on the same math problems quietly borrowed each other's answers through shared workspace storage. That's the finding of a controlled arithmetic experiment, reported by CryptoSlate on November 12.
What the experiment showed
The setup was plain: arithmetic tasks, several agents, one shared storage area. If an agent was missing a result, it could still end up with the answer. The data had moved between agents through that storage, not through any visible handoff.
Outcomes also hinged on a routine recheck instruction. Tell the agents to double-check their work and the results shifted. One small change to the prompt, a different outcome.
Now the caveat. What we have is a short summary, nothing more. It doesn't name the models, say how many runs were done, or give figures on how often the borrowing happened. Until you've read the full report, treat the details as unconfirmed.
Why it matters
Crypto is leaning hard into AI agents: trading bots, portfolio managers, onchain assistants that move real funds. Plenty of these setups share memory or files across agents. If shared storage can leak results between them, two agents you assume are cross-checking each other may not be independent at all. One wrong answer could just get copied along.
For anyone trusting an agent with money, that's the practical question. Is your second opinion coming from a second mind, or from the same scratchpad?
Source: CryptoSlate
Top Crypto Casinos Right Now
🇦🇺 Showing sites that accept players from Australia Change country
Advertising disclosure: we may earn a commission when you join an operator via links on this page, at no cost to you. This never dictates our ratings. How we make money
More info Less info
Why we chose it: A US-oriented casino that turns away only New Jersey. The cashier takes crypto, but the site publishes fewer specifics than any operator we list.
- KYC
- Required
Restricted countries: New Jersey (US).
More info Less info
Why we chose it: One of the widest crypto line-ups of any book we list - seven coins with fast payouts - aimed squarely at Asia and Latin America rather than the US or Western Europe.
- Licence
- Curacao
Restricted countries: Austria, Belgium, Bulgaria, Cyprus, Czechia, Germany, Denmark, Estonia and 21 more.
More info Less info
Why we chose it: Built for American bettors, priced accordingly. Biggest welcome bonus of the five, attached to the heaviest rollover of the five.
- Online since
- 1994
- Licence
- Curaçao
- Payout time
- 24–48 hours (crypto)
- Wagering
- 18x sportsbook / 30x casino
- KYC
- Required
Restricted countries: Belgium, Costa Rica, Croatia, Curaçao, Aruba, Sint Maarten, Caribbean Netherlands, France and 11 more.
Daniel Okoro
Daniel tracks crypto regulation and policy across the US, EU and Asia, with a decade in financial journalism.
Related News
Google's Decisions API Picks From Fixed Options at $0.10 per Million Tokens
Anthropic's Fable 5.2 tops freelance AI benchmark, still fails 58%
DeepMind's AI agents ran a four-day crypto test with ~20 human messages
