An AI Found a Bug That Could Mint 18 Trillion XRP. The Reward Was $250,000.

On September 21, an AI agent built by a small security startup called Veria Labs was pointed at `rippled`, the software that runs the XRP Ledger. By the next day it had found two integer overflows, worked out how to chain them, and built a working exploit on a local network. According to Veria's own write-up, a single transaction could have created about 18.45 trillion XRP, roughly 184 times the token's fixed 100 billion supply. The code behind it dated to 2015.

RippleX confirmed the report the same day, merged a fix on September 23 and shipped it in xrpld 3.4.1 on September 25, with more than 80% of default validators upgraded by release day, The Block reported. There is no evidence anyone exploited it. On October 8, Veria was paid the program's maximum bounty: $250,000.

A quarter of a million dollars. For the integrity of a $94 billion asset.

That works out to about 0.00027% of the value Veria estimated was at risk. On Sunday, the on-chain investigator ZachXBT called the payout cheap and said the finding "should be a 7 fig pay day at minimum." He is right about the number, but the bigger problem is the logic behind it. The industry still prices bug bounties like a budget line, while attackers price exploits by what they can take. AI just made finding these bugs much cheaper for both sides.

The price of honesty in crypto is now being set by the people who steal.

The way the bug works matters, because it explains why it went unnoticed for a decade. The payment engine adds up what a buyer owes when one payment fills many order book offers at once, and that sum had no overflow check. Push the total past the limit of a 64-bit integer and it wraps around to a tiny number. The buyer pays a few hundred drops, and every seller is still credited in full. A separate safety check, the invariant that exists specifically to reject any transaction that creates XRP, had the same flaw. Veria's agent set up 256 offers so that both counters wrapped by exactly the same amount, and the check saw what looked like an ordinary fee burn.

Each overflow on its own was low severity. Chained together, they broke the one promise XRP has always made about its supply. Veria said general-purpose coding agents pointed at the same code afterward "still couldn't find it."

The bounty market was built for a slower kind of attacker.

The standard reward formulas in crypto security assume finding a critical bug is rare, skilled, slow work. Immunefi's default program, a common template, pays 10% of the funds directly affected, capped at $50,000 for a critical smart contract bug. Projects can set higher ceilings, and the famous ones did. Polygon once paid $2 million for a bridge flaw that put about $850 million at risk, roughly 0.24% of the exposure.

Zcash is the most recent comparison, and the one ZachXBT used. Security researcher Taylor Hornby, also working with AI assistance, found a counterfeiting bug in Zcash's Orchard shielded pool that put about $2.3 billion of ZEC at risk. His original $750,000 award was about 0.033% of that, a figure Shielded Labs' Jason McGee called too low before nominating a matching $750,000 bonus that coinholders approved. McGee described the extra money as "cheap insurance."

Put those ratios next to each other. Polygon paid about a quarter of a percent. Zcash's holders decided a thirtieth of a percent was not enough and doubled it. The XRP Ledger paid less than a thousandth of a percent for a bug that threatened the whole market cap.

Attackers already have their own price list.

The other side of the market posts its prices too. On September 6, an attacker exploited a validation flaw in Elements, the software behind Blockstream's Liquid Network, minted about 4,000 unbacked L-BTC and had roughly $319 million in real bitcoin paid out 36 minutes later. The attackers wrote "we are whitehats" on-chain, negotiated with Blockstream, and sent back about 85% once the bug was patched. They kept 598.5 BTC, about $47 million, which TRM Labs says they appear to be claiming as a bounty.

That is the real benchmark. A team that reported the Liquid bug politely would have been paid whatever Blockstream's program allowed. A team that exploited it first and negotiated afterward walked away with about 188 times what Veria received for the XRP bug. The incentive could not be clearer, and nobody had to write it into a policy.

The scale of losses shows how many people already understand that. Finbold, using SlowMist data, counted more than $1.17 billion stolen in the third quarter across about 125 incidents, and 11 more hacks in the first nine days of October.

AI has collapsed the cost of finding bugs, but not the cost of paying for them.

The Veria episode is a preview. A machine found a ten-year-old bug in two days that years of human review had missed. Veria's write-up also points to this summer's Coldcard firmware exploit, in which 1,367 BTC was drained from wallets after a 2021 entropy flaw, as a case Coinkite suspects involved AI. The same tools are available to everyone. When the cost of searching a codebase falls toward the cost of compute, more people search, and more of them face the choice between reporting and exploiting.

Bounty programs were designed for a world where those people were a small, specialized group whose reputations kept them honest. That group is getting larger and less tied to anyone's reputation. The only thing left to keep them honest is the price.

Ripple's best defense is real, and it still falls short.

There is a fair defense, and it should be stated plainly. The $250,000 was the published maximum, and Veria knew the terms when it reported. The fix shipped in three days. That was fast enough that the network skipped its usual two-week amendment vote for the first time in more than a decade, which says the team took the threat seriously. And "$94 billion at risk" overstates what any attacker could actually have taken. Nobody can sell 18 trillion XRP. The order books are not that deep, exchanges would freeze deposits within minutes, and validators have shown they can coordinate quickly. Q3 had its own example of this: the Tectonic attackers moved less than $10 million before that chain was halted and rolled back.

All of that is true, and the thesis still holds. The realistic payoff from an infinite-mint bug is not the market cap. It is whatever an attacker can push through exchanges, bridges and decentralized order books before anyone notices, plus whatever they can win in a negotiation afterward. Liquid showed that the first part can mean hundreds of millions of dollars in well under an hour, and that the second part can mean keeping $47 million with a white-hat label attached. Neither number is anywhere near $250,000.

There is also a cost that never appears in a bounty budget. XRP's fixed supply is not one feature among many. It is the asset's basic claim. One successful mint, even one that was rolled back, would have turned "capped at 100 billion" into "capped unless someone finds the next overflow." That is a much more expensive outcome than a seven-figure check.

Credit to the XRP Ledger Foundation contributor who replied to ZachXBT that the program is being strengthened and that "we have to make sure the incentive structure stays attractive with the value at risk," as Bitcoin.com reported. RippleX's head of engineering has also said the team will expand AI-assisted bug hunting and formal verification. Both are the right moves. Neither changes the fact that the next agent to find a bug like this may belong to someone who never reads the bounty terms.

Every protocol now has two prices on its worst bug: the one it publishes, and the one an attacker can collect. The protocols that last will be the ones that close the gap before an AI shows them how big it is.