22h ago

Aman Priyanshu finds LLM agents are eight times more likely to leak private data after observing another agent overshare

The leaks occur naturally without external adversarial attacks

——0——
Original post
Niloofar#303@NILOOFAR_MIREOPAman Priyanshu @ ACM CAIS 2026Aman Priyanshu @ ACM CAIS 2026@AMANPRIYANSHU6

What does an AI agent do with your private data the moment it sees another agent overshare theirs? Spill yours. About 8x more likely to. And nothing has to attack it. For 25 simulated days we ran 2,533 LLM agents on a Moltbook-style platform with synthetic human profiles attached, then ran 7,000 controlled trials across 7 frontier models. Three findings:

1:07 PM · May 28, 2026 View on X
Reposted by
Seth Lazar#1060@SETHLAZAR
QUOTE POSTNiloofar#303Niloofar@NILOOFAR_MIRE

Super cool work on simulating moltbook agents for long terms (tens of days) and studying their secret keeping behaviors!

Aman Priyanshu @ ACM CAIS 2026Aman Priyanshu @ ACM CAIS 2026@AmanPriyanshu6

What does an AI agent do with your private data the moment it sees another agent overshare theirs? Spill yours. About 8x more likely to. And nothing has to attack it. For 25 simulated days we ran 2,533 LLM agents on a Moltbook-style platform with synthetic human profiles attached, then ran 7,000 controlled trials across 7 frontier models. Three findings:

8:07 PM · May 28, 2026 · 8.4K Views
8:24 PM · May 28, 2026 · 5.8K Views