This isn't a hypothetical risk

In 2018, a privacy researcher downloaded 208 million public Venmo transactions. Buried in the payment notes, the ones people assumed were harmless, were drug deals, breakups, and daily routines. A later academic study at USC went further, examining 389 million of those notes. One in ten leaked something genuinely sensitive: a health condition, a political view, a habit someone never meant to share.

Nobody thought a note for coffee or rent was a privacy risk. It was. AI agents now make far more payments than any person ever will, often hundreds a day, with nobody reading each one before it goes out. Many of those payments settle on public blockchains, so anything caught in the metadata isn't just at risk of a future leak. It can be permanently and publicly readable from the moment the payment happens.

PennyWall exists because this pattern is easy to predict and hasn't been solved for agents yet. It catches what's in a payment before it leaves, not after somebody downloads a public feed and writes about it.

Sources: Hang Do Thi Duc, "Public by Default" (2018) · USC Information Sciences Institute, analysis of Venmo transaction data (2022)