AIDA: An AI Outage Registry with Evidence from On-Chain MEV
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Quantitative Finance. The day in quantitative finance: models of markets, risk and prices.
Ruth: We're looking at the paper: "AIDA: An AI Outage Registry with Evidence from On-Chain MEV".
Warren: Here's the gist of the paper. Generative AI has made trading software cheaper to write, but no systematic dataset has recorded when these services fail.
Ruth: First, entry into MEV expanded where code and public state suffice, as active arbitrage addresses quadrupled from late 2024 to late 2026 and liquidators grew by a third, while sandwich attackers, who need private order flow and builder access, declined.
Warren: Second, AI outages stall new trading code while deployed bots keep trading across both chains.
Ruth: Data Construction and Identification The study introduces the AI Incident and Disruption Archive (AIDA), a multi-vendor registry of 5,868 deduplicated outage incidents across 30 providers from 2025 to 2026, with foundation-model incidents verified against primary status-page or postmortem URLs.
Paper summary: Ruth: So we’re talking about this paper today from arXiv, titled "AIDA: An AI Outage Registry with Evidence from On-Chain MEV". It looks at how generative AI is changing the landscape for writing trading software and what happens when that technology goes down.
Warren: Yeah, so the core idea is that even though writing trading code has gotten cheaper because of these AI tools, nobody has a systematic dataset to tell us when those services actually fail. That’s what this paper claims they’ve created.
Iris: It sets up this thing called AIDA, which is a multi-vendor registry of five thousand eight hundred sixty-eight deduplicated outage incidents across thirty providers from two thousand twenty-five to two thousand twenty-six and they verified the foundation model incidents against status pages or postmortems > <ref:2611.01656#pg1,a multi-vendor registry of 5,868 deduplicated outage incidents across 30>
Ruth: And they paired that with about five million verified maximal extractable value transactions by eight thousand five hundred sixteen Ethereum addresses, contract creations, and a two-two month panel on Base Layer two with over one billion swaps >
Warren: They’re looking at atomic arbitrages, sandwich attacks, and lending liquidations from primary rules across the whole Ethereum span over five million verified strategy transactions by eight thousand five hundred sixteen addresses > <ref:2611.01656#pg1>
Iris: The main point is they establish two findings based on this data >
Ruth: Okay, so what are those two things they found when you look at this AIDA registry and that transaction data?
Warren: First, entry into maximal extractable value expanded exactly where writing code and public state were enough, which makes sense since the cheaper software lowers the barrier to entry >
Iris: They found that active arbitrage addresses grew four times between late two thousand twenty-four and late two thousand twenty-six and liquidators grew by a third, while sandwich attackers actually shrank > <ref:2611.01656#pg1,late 2026, and liquidators grew by a third, while sandwich attackers>
Ruth: So it seems like AI is making it easier for people to build things that rely on code and public state, but the more complex stuff requiring private order flow isn't as much affected right now >
Warren: Right. And second, they found that AI outages stall the creation of new trading code while bots keep trading across both chains >
Iris: On Ethereum, addresses releasing a new executor contract dropped sharply on the day of a multi-hour disruption and then rebounded within days, which suggests a multi-day development pipeline >
Ruth: So what this means for someone listening right now is that developers probably use AI for the off-chain engineering work, but when those AI services stop working, the live execution bots just keep trading without missing a beat >
Warren: It points to how they rely on these coding services for building things, but the actual live trades run autonomously when the systems fail >
Iris: The study notes that private order flow and latency still determine who wins execution even though cheaper code is making it easier to build those contracts >
Ruth: So the implication here is that AI helps with off-chain engineering while the on-chain trading part keeps running on its own, but you still need good execution speed and capital to win in the actual market >
Warren: Exactly. The paper's title, "AIDA: An AI Outage Registry with Evidence from On-Chain MEV", frames it as a way to track this whole relationship between AI coding tools and live on-chain trading >
Iris: It really tries to resolve an identification problem that price and volume studies struggle with because aggregate data alone doesn't tell you if a disruption is affecting the building of software or the execution of live trades >
Ruth: So what do we get from this paper in terms of what it changes for us as market observers?
Warren: It gives us evidence that AI coding tools are helping MEV developers with off-chain engineering, but those developers still depend on live execution running autonomously when the AI services go offline >
Iris: The study shows that the entry pattern confirms this: the expanding base of MEV developers enters exactly where code and public state suffice because of cheaper software >
Ruth: It seems like it’s a two-part story, then: AI helps build things faster, but the live execution itself is more resilient than we thought as long as those bots are deployed >
Warren: That resilience is tied to how long the development pipeline takes, which the outage data shows can be multi-day when services are down >
Iris: It’s important to remember that the researchers are trying to separate whether a disruption affects building trading software or live trades, and their findings show they're complementary >
Ruth: So if you only listen to this paper, the big picture is that we need to look at both the development side and the execution side when we analyze how AI impacts MEV >
Conclusion: Ruth: So we’re wrapping up this look at AIDA, which is this AI Outage Registry with Evidence from On-Chain MEV paper by those researchers on arXiv today.
Warren: Basically, they built a way to track when AI coding services actually fail while you're trading on the blockchain.
Iris: They used this archive of outage incidents and paired it with massive amounts of data, like millions of swaps and strategy transactions across Ethereum and Base Layer two.
Ruth: The authors are trying to untangle whether these outages stop new software from being built or if the bots that are already running can keep trading without issue.
Warren: It looks like they found that AI helps developers write off-chain code, but when those services drop, the live execution bots just keep running across both chains.
Iris: The data shows a sharp drop in new executor contracts on Ethereum right after a disruption, which suggests developers rely on those AI tools for engineering but the live trading part is pretty resilient.
Ruth: So what this means for us is that we need to look at both the development side and the execution side when we analyze how AI impacts MEV.
Warren: It means that even if you have cheap code, who wins execution still comes down to who has fast access and enough capital because of latency and order flow.
Iris: And this paper shows how important it is to separate the building part from the live trading part when you’re trying to understand this whole system.
Wenbin Wu
Cambridge Centre for Alternative Finance, Cambridge Judge Business School, University of Cambridge
q-fin.TR, cs.CE
Submitted: 2026-11-01
Updated: 2026-11-04
Comments: 27 pages, 7 figures, 5 tables
Code: https://github.com/dthinkr/agentic-mev
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 100/100
The gist: Generative AI has made trading software cheaper to write, but no systematic dataset has recorded when these services fail.
Key concepts
- AI Incident and Disruption Archive (AIDA)
- This is a registry tracking 5,868 deduplicated outage incidents from 30 providers between 2025 and 2026. It verifies foundation model outages against official status pages or postmortems to provide concrete data on when AI services fail.
- Maximal Extractable Value (MEV) Transactions
- These are verified transactions that represent the highest possible value extracted from blockchain activity. The study uses these to track specific trading strategies, such as atomic arbitrages and sandwich attacks, across Ethereum and Layer-2 chains.
- Run-Time Dependence
- This refers to how trading code behaves when AI services are unavailable. The research shows that while new code development stops during outages, existing deployed bots continue to execute trades autonomously on the blockchain.
- Sandwich Attackers
- These are traders who exploit transaction ordering by placing one transaction before and another after a target transaction to profit from the price movement. The study noted that these attackers declined as AI made building such specialized code more difficult.
Terminology
Summary
Generative AI has made trading software cheaper to write, but no systematic dataset has recorded when these services fail. p. 0 AIDA: An AI Outage Registry with Evidence from On-Chain MEV
Wenbin Wu
Cambridge Centre for Alternative Finance, Cambridge Judge Business School, University of Cambridge
25 Trumpington Street, Cambridge CB2 1QA, United Kingdom
Abstract
Generative AI coding tools have lowered the cost of writing trading software, but no systematic dataset records
when these services fail. We introduce the AI Incident and Disruption Archive (AIDA), a multi-vendor registry of
5,868 deduplicated outage incidents across 30 providers from 2025 to 2026, with foundation-model incidents verified
against primary status-page or postmortem URLs. Pairing AIDA with 5,041,034 verified maximal extractable value
(MEV) transactions by 8,516 Ethereum addresses, their contract creations, and a 22-month Base Layer-2 panel of
1,099,593,191 swaps and 29,120 private executors, we establish two findings. First, entry into MEV expanded where
code and public state suffice, as active arbitrage addresses quadrupled from late 2024 to late 2026 and liquidators grew
by a third, while sandwich attackers, who need private order flow and builder access, declined. Execution nonetheless
remained concentrated, with the top fifty addresses executing two thirds of verified trades. Second, AI outages stall
new trading code while deployed bots keep trading.Where the paper says it
Where the paper says it
Where the paper says it
Data Construction and Identification
The study introduces the AI Incident and Disruption Archive (AIDA), a multi-vendor registry of 5,868 deduplicated outage incidents across 30 providers from 2025 to 2026, with foundation-model incidents verified against primary status-page or postmortem URLs. This archive is paired with verified maximal extractable value (MEV) transactions by 8,516 Ethereum addresses, their contract creations, and a 22-month Base Layer-2 panel of 1,099,593,191 swaps and 29,120 private executors. p. 0 We introduce the AI Incident and Disruption Archive (AIDA), a multi-vendor registry of
5,868 deduplicated outage incidents across 30 providers from 2025 to 2026, with foundation-model incidents verified
against primary status-page or postmortem URLs. Pairing AIDA with 5,041,034 verified maximal extractable value
(MEV) transactions by 8,516 Ethereum addresses, their contract creations, and a 22-month Base Layer-2 panel of
1,099,593,191 swaps and 29,120 private executors, we establish two findings. Answering these questions requires transaction-level strategy identification and a systematic record of AI service
disruptions. We identify atomic arbitrages, sandwich attacks and lending liquidations from primary-source rules over
raw event logs, favoring precision over recall, across the full Ethereum span (5,041,034 verified strategy transactions by
8,516 addresses; Table 1) and a 22-month Base panel.Where the paper says it
Where the paper says it
Findings on Strategy Entry and Execution
The entry pattern shows that the expanding base of MEV developers enters exactly the strategies where code and public state suffice, consistent with cheaper software lowering the building barrier. Specifically, active arbitrage and liquidation addresses expanded markedly between the fourth quarter of 2024 and the third quarter of 2026, while sandwich attackers contracted (Table 2). p. 1 First, entry into MEV expanded where code and public state suffice. Active arbitrage and
liquidation addresses expanded markedly between the fourth quarter of 2024 and the third quarter of 2026, while
sandwich attackers, who need private order flow and builder access, contracted (Table 2). First, entry into MEV expanded where
code and public state suffice, as active arbitrage addresses quadrupled from late 2024 to late 2026 and liquidators grew
by a third, while sandwich attackers, who need private order flow and builder access, declined. Execution nonetheless
remained concentrated, with the top fifty addresses executing two thirds of verified trades.Where the paper says it
Where the paper says it
Findings on Run-Time Dependence
AI outages stall new trading code while deployed bots keep trading. On Ethereum, addresses releasing a new executor contract decline sharply on the day of a multi-hour disruption and rebound within days, consistent with a multi-day development pipeline. p. 2 Second, AI outages stall new trading code while deployed bots keep trading. On Ethereum, addresses releasing
a new executor contract decline sharply on the day of a multi-hour disruption and rebound within days, consistent
with a multi-day development pipeline (Table 3). The outage design then provides direct evidence that these developers rely on AI coding services, because
when those services fail, new executor releases pause while the same operators' deployed bots trade on uninterrupted. Together, the findings establish that AI assists MEV in off-chain software engineering while live on-chain execution
runs autonomously.Where the paper says it
Where the paper says it
Market Context and Limitations
The findings establish that AI assists MEV in off-chain software engineering while live on-chain execution runs autonomously. The study notes that private order flow, latency and capital still determine who wins execution, even as cheaper code lowers the barrier to building MEV contracts. p. 0 Cheaper code thus lowers the barrier to building MEV
contracts, while private order flow, latency and capital still determine who wins execution. Keywords: decentralized exchanges, arbitrage, maximal extractable value, generative AI, outages, event study
JEL classification: G14, G23, O33
1. Some operators execute atomic trades inside a contract's constructor without subsequent
calls, so that the contract-creation transaction itself constitutes a trade execution, many deploy token, helper or test
contracts unrelated to MEV, and on low-gas Layer-2 rollups such as Base existing contracts also spawn large numbers
of internal factory clones automatically during routine trading. We therefore classify Ethereum creations by the created
contract's subsequent role into persistent private executors that go on to execute identified MEV transactions, broader
swap-receiving trading contracts, single-use constructor-executed trades, and non-trading contracts, while on Base
we also separate top-level creations signed directly by an operator key, distinct operators deploying an executor, and
per-operator capped executor creations from uncapped internal factory spawns.Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Develop a system that separates off-chain development from on-chain execution by tracking
AI outages
via AIDA, allowing it to determine ifnew trading code stalls while deployed bots keep trading.
This enables precise measurement of whetherAI services disrupt new trading code while deployed bots keep trading.
-
Implement a strategy identification module that classifies transactions into atomic arbitrage, sandwich attacks, and lending liquidations based on transaction-level rules, rather than relying on coarser metrics like pool count or block positions. This allows the system to distinguish between strategies requiring only
code and public state
versus those needingprivate order flow and builder access.
-
Create a deployment monitoring tool that tracks the release of new executor contracts by correlating them with AIDA incident data, specifically measuring the effect of disruptions on these releases, as indicated by results showing that
addresses releasing a new executor contract decline sharply on the day of a multi-hour disruption.
-
Design an execution monitoring layer that assesses live trading performance during AI outages, confirming whether
live arbitrage and sandwich execution across both chains remain unchanged
when compared to reference rates during disruption windows. -
Build an address-level deployment classifier that distinguishes between
persistent private executors,
broader swap-receiving trading contracts,
andnon-trading contracts
to accurately measure the development margin under AI service disruptions, as this separation is crucial for validating the hypothesis thatMEV developers rely on AI coding assistants when writing, testing and shipping new trading contracts.
Abstract
Generative AI coding tools have lowered the cost of writing trading software, but no systematic dataset records when these services fail. We introduce the AI Incident and Disruption Archive (AIDA), a multi-vendor registry of 5,868 deduplicated outage incidents across 30 providers from 2025 to 2026, with foundation-model incidents verified against primary status-page or postmortem URLs. Pairing AIDA with 5,041,034 verified maximal extractable value (MEV) transactions by 8,516 Ethereum addresses, their contract creations, and a 22-month Base Layer-2 panel of 1,099,593,191 swaps and 29,120 private executors, we establish two findings. First, entry into MEV expanded where code and public state suffice, as active arbitrage addresses quadrupled from late 2024 to late 2026 and liquidators grew by a third, while sandwich attackers, who need private order flow and builder access, declined. Execution nonetheless remained concentrated, with the top fifty addresses executing two thirds of verified trades. Second, AI outages stall new trading code while deployed bots keep trading. On Ethereum, addresses releasing a private executor fall to 0.67 on the day of an eight-hour disruption (0.68 for executor creations during OpenAI outage hours) and to 0.66 at lag two before rebounding on day three; on Base, operator top-level creations fall to 0.76 during OpenAI hours and executor-deploying operators to 0.88 at lag two. Non-trading creations, automated factory clones, and live arbitrage and sandwich execution across both chains remain unchanged. Cheaper code thus lowers the barrier to building MEV contracts, while private order flow, latency and capital still determine who wins execution.
Sources
- Remeasuring the Arbitrage and Sandwich Attacks of Maximal Extractable Value in Ethereum
- The Impact of AI on Developer Productivity: Evidence from GitHub Copilot
- EVAGE: Autonomous MEV Generation and Adaptation via Multi-Agent Harness
Related papers
- MLP-Initialized Reinforcement Learning for Exit Timing in SPX Short-Straddle Portfolios
- Beyond Anonymous Order Flow: Wallet Identity as a Feature for Short-Horizon Return Prediction
- Combinatorial Arbitrage in Prediction Markets
- TradeMech: A Method to Multilaterally Net Trades Without Altering Counterparty Exposure
- Optimal Trading Speed on Blockchain
- Why financial prices exhibit Brownian motion despite predictable order flow: An exactly solvable nonlinear price impact model