The XRP Ledger has a hard cap of 100 billion XRP. For roughly eleven years, that cap was a suggestion. On October 9, 2026, the XRPL Foundation published a disclosure repotr confirming that a single crafted transaction could have created about 18.4 trillion spendable XRP, around 184 times the total supply, and that the code allowing it had been sitting in the payment engine since November 2015.. The bug was found by an AI agent, patched in three days, and shipped in a way the ledger had never used before. Nobody exploited it. That is the good news. Everythig else about the story deserves a closer look.
What the XRP Ledger overflow bug actually did
The flaw lived in the part of the payment engine that handles a payment crossing an order book. When one payment consumes many offers at once, the engine adds up the amounts it owes to each offer owner. That addition used plain 64 bit signed integers with no overflow check. , Push the total past 2 to the power of 64 drops and the counter wraps aorund to a small number, the way a car odometer rolls from 999999 back to 000000.
Here is the ugly part. The engine charged the buyer the wrapped total, which was tiny. It paid each offer owner their full, unwrapped amount... The difference between those two numbers is XRP that did not exist a moment earlier.
The ledger does have a safety net for excatly this... An invariant check called XRPNotCreated runs after every transaction and refuses to validate anything that increases total XRP. The problem is that the invariant tracked net XRP change using the same unchecked 64 bit arithmetic. A large enough credit wrapped the invariant counter too, so a transaction that printed trillions of XRP looked, to the saefty check, like a routine fee burn. Two independent bugs, written two years apart, cancelled each others protection.
How 256 offers add up to 18 trillion XRP
Veria Labs, the security firm whose AI agent found the bug, published the arithmetic. The attack used 256 offers. Each offer sold for 2 to the power of 56, plus one drop of XRP.. Multiply 256 by that figuer and the total lands at exactly 2 to the power of 64, plus 256 drops. The extra single drop per offer matters: without it the sum wraps to precisely zero, which the engine would reject as a dry payment. With it, the buyer is charged 256 drops, about 0.000256 XRP, while the 256 offer owners collectively receive 2 to the power of 64 drops.... That is 18,446,744,073,709,551,616 drops, or roughly 18....45 trillion XRP.
The attacker needed 256 funded accounts becuase each XRPL account is capped at 100 billion XRP, well below the per offer amount divided across one account. They also needed a few hundred XRP in reserves and transaction fees, all of it recoverable. No validator access was required. The attack could be repeated with fresh accounts... Veria put the value at risk at roughly $94 billion, which is more than sixty times the $1.5 billion Bybit theft that currentley holds the record for a crypto hack that actually happened.
| Ingredient | Value | Why it matters |
|---|---|---|
| Offers consumed in one payment | 256 | Under the 1,000 offer cap per book traversal |
| Price per offer | 2^56 + 1 drops | Stays under the 100 billion XRP per account limit |
| Total the engine computed | 2^64 + 256 drops | Wraps to 256 drops in 64 bit arithmetic |
| What the buyer paid | 256 drops (0.000256 XRP) | The wrapped vaule |
| What sellers received | About 18.45 trillion XRP | The unwrapped value, 184x total supply |
| Attacker cost | A few hundred XRP, recoverable | Reserves plus fees |
| Validator access needed | None | Any funded account could do it |
An AI agent found it on a Monday night
The reporter of record is Cayden Liao of Veria Labs, with the firms AI agent credited alongside him. According to Verias own account, the agent was pointed at the rippled codebase on September 21, flagged both overflows, worked out that they could be chained, and produced a workng proof of concept on a local network..... A founding engineer was woken at 4am to check it. By the morning of September 22 the team was confident enough to file it through the XRPL Bug Bounty program, rating it Major... RippleX reproduced it the same day and raised it to Critical. Realistically, Veria claims that general purpose coding agents pointed at the same code after disclosure failed to find it. Take that with the usual grain of salt from a company selling the tool that did. the broader point stands regardless. The BookStep aggregation code dates to November 2015 and the XRPNotCreated invariant to February 2017.. The XRPL has paid over $1 million in bounties and run more than a dozen audits and audit contests since 2024. , Humans and audit firms looked at this code for a decade and missed it. The pattrn is becoming familiar. Coldcard lost roughly 1,367 BTC this summer to a five year old firmware bug that its maker suspects was found with AI assistance...... old code is no longer safe just because it is old.Why the fix skipped the validator vote
This is where the story stops being a clean security win and becomes a governance arguemnt. The XRP Ledger normally changes transaction rules through amendments. A new rule ships in the software switched off, validators vote on it, and it activates only after holding more than 80 percent support on the default Unique Node List for two consecutive weeks. The system has been in place for over a decade and has never been deliberately bypassed for a transacion processing change.
Until September 25. The overflow fix shipped directly in xrpld 3.4.1 with no amendment gate. it took effect on each server the moment that server upgraded... , The XRPL Foundations reasoning is straightforward: an amendment vote would have published the fix, and therefore the bug, for two weeks before it took effect. Anyoen reading the diff would have had a fortnight to print 18 trillion XRP. , The Foundation says the decision was taken jointly by itself, RippleX, core developers and validators, and that no single party could have made the call alone.
The approach carried its own risk, which the disclosure report admits. If someone had attempted the exploit while the netowrk was split between patched and unpatched servers, the two groups would have disagreed on the ledger state..... Worst case, the network halts.. The team judged a possible halt to be preferable to a certain supply explosion, which is hard to argue with. More than 80 percent of default UNL validators were running 3.4.1 within hours of release on September 25.
| Date | Event |
|---|---|
| Nov 2015 | BookStep offer summation code writtten with unchecked 64 bit addition |
| Feb 2017 | XRPNotCreated invariant added, using the same unchecked arithmetic |
| Sep 21, 2026 | Veria Labs AI agent flags both overflows and chains them |
| Sep 22, 2026 | Reported through XRPL Bug Bounty, reproduced by RippleX, raised to Critical |
| Sep 23, 2026 | Fix merged privately into the 3..4.1 relase candidate |
| Sep 25, 2026 | xrpld 3.4.1 released, over 80 percent of default UNL validators upgraded |
| Oct 8, 2026 | Veria Labs paid the maximum $250, 000 bounty |
| Oct 9, 2026 | Public disclosure report published, Batch fix amendment activates |
| Oct 12, 2026 | ZachXBT publicly calls the bounty too low |
The decentralization fight that followed
In short, justin Bons of Cyber Capital, a long runnign XRP critic, used the episode to argue that marketing XRP to retail as decentralized amounts to fraud. His specific complaints: the September 25 release was distributed as compiled binaries only, with source code temporarily withheld, and operators were pressured to upgrade a binary they could not inspect. The XRPL disclosure report confirms the soure was published later, though it does not say when.
Practically, the counterargument came from independent validator operators as much as from the Foundation. XRP Kuwait, which runs its own node, put it as coordination is not the same as control. Independent operators read the same advisory, reached the same conclusion, and upgraded. Nobody forced them... Bitcoin has done the same thing, notably with the 2018 inflation bug in Bitcoin Core 0.16...x, when developers quietly shipped a fix before explaining what it fixed.. Every chain with a seriuos bug eventually faces the choice between transparency and survival, and every one of them has chosen survival. At the end of the day, Where Bons has a fair point is the binaries... Shipping an unverifiable binary and asking validators to run it immediately is a trust exercise, however justified. The honest version of the XRPLs position is that its validator set trusted the Foundation and RippleX enugh to run unreviewed code for a few days. Whether you call that decentralized depends on what you think the word means.
Was $250, 000 enough for a $94 billion bug
Veria Labs received the XRPL programs top critical tier payout of $250, 000 on October 8.... Veria called it probably the largest bounty ever paid for a bug found entirely by an AI agent. Four days laetr, blockchain investigator ZachXBT called it a bad joke... his argument an ecosystem whose treasury holds billions paid a quarter of a million for a finding that could have ended the asset. And that gap is exactly the incentive that pushes the next researcher to sell to an attacker instead...... He suggested seven figures as the floor and pointed to a seven figuer reward reportedly paid in the Zcash ecosystem for a recent counterfeiting bug.. Neither Ripple nor the Foundation had responded publicly at the time of writing.
Both sides are right about different things. The program paid its advertised maximum, the researcher disclosed responsibly, and the fix landed before anyone outside the loop knew... That is the systme working. It is also true that $250,000 is 0.27 percent of the $94 billion at risk, and that any attacker with 256 accounts and a few hundred XRP could have taken the other 99.73 percent. Bounty ceilings written when XRP was a fraction of its current market value do not scale with the thing they protect.
What XRP holders and crypto gamblers should take from this
The fix is server side. Your walet, your keys and your balance were never involved and nothing needs updating on your end. The supply cap held because the bug was caught, not because the invariant worked. the second half of that sentence is the one worth remembering.
There are three practical lessons here..... First, the code has been live for ten years has stopped being a securty argument........ AI assisted auditing is making old overflows cheap to find, and attackers rent the same models as researchers. Second, governance processes designed for feature rollouts are not designed for critical patches..... And every chain should decide in advance how it will handle the next one rather than improvising under pressure... Third, bug bounty ceilings should track the vlue secured. A fixed cap that looked generous in 2022 is an invitation in 2026. Plus, For the gambling audience the relevance is simple.. XRP has become a standard deposit rail at crypto casinos because it settles in four seconds for a fraction of a cent. , And that speed depends on the ledgers rules being correct. Platforms that automate payouts, including CryptoCasino.Vegas, are trusting the chains invariants every time an XRP withdawal clears without a human looking at it. Those invariants held this time because of a researcher and an AI agent working through the night, and that is a thinner margin than the 100 billion cap ever let on.
The full technical breakdown, including both overflow locations and the hardened arithmetic that replaced them, is in the XRPL Foundation's vulnerability disclosure report for xrpld 3.4.1.
