h hoge.gg
Subscribe
BTC$67,432.18+2.34%ETH$3,521.44+1.08%SOL$178.62-0.62%BNB$612.30+0.41%XRP$0.6234-0.18%ADA$0.4521+3.12%DOGE$0.1623+1.86%AVAX$38.71-1.24%LINK$17.84+0.92%HOGE$0.00004120+4.21%
BTC$67,432.18+2.34%ETH$3,521.44+1.08%SOL$178.62-0.62%BNB$612.30+0.41%XRP$0.6234-0.18%ADA$0.4521+3.12%DOGE$0.1623+1.86%AVAX$38.71-1.24%LINK$17.84+0.92%HOGE$0.00004120+4.21%
● AI x Crypto

AI Agents On-Chain: Who Pays When the Agent Acts in 2026

On November 24, Wells Fargo makes customers liable for whatever their AI agent does, even transactions it was tricked into. Here is who actually pays when an autonomous agent moves your money.

On November 24, 2026, one of the largest banks in the United States will quietly change who is responsible when a piece of software spends your money. Wells Fargo’s updated Online Access Agreement adds a new Section 15 covering third-party AI agents, and its logic is blunt: if you hand an agent your login, anything it does is treated as if you did it yourself. The bank does not have to verify the instruction, does not monitor the agent, and tells you in capital letters that using one is at your own risk.

Read the clause slowly and the implication lands. Section 15 defines an AI agent as any software that takes autonomous, semi-autonomous, or non-autonomous action on your behalf. If that software signs a transaction you never intended, or acts on an input an attacker slipped into its context window, the loss is yours. You even have to keep proof that you revoked the agent’s access, because you stay on the hook until you tell the bank. This is not a thought experiment about some distant machine economy. It is a liability regime, written in plain contract language, shipping to retail customers this week.

Wells Fargo is not an outlier. Between October 13 and 20, JPMorgan Chase, Citigroup, and Charles Schwab each moved to block or contractually reallocate the risk of customer-directed agents logging in with the customer’s own credentials, with one legal summary noting that JPMorgan chief executive Jamie Dimon warned about “rogue agents” on the bank’s October earnings call and the “liability transfer” that comes with letting them in. The question this article asks is the one the banks are answering for themselves and leaving everyone else to work out: when an autonomous agent acts on-chain and something goes wrong, who actually pays?

What an On-Chain AI Agent Actually Is

Strip away the hype and an on-chain AI agent is three things bolted together: a large language model that plans and interprets, a wallet it controls, and a set of tools it can call. The model reads a goal in plain language, decides what to do, and then signs its own transactions. That last part is what separates an agent from the trading bots crypto has run for a decade. A bot executes a fixed rule you wrote. An agent writes the rule on the fly, which is exactly why it is useful and exactly why it is dangerous.

The autonomy is the whole point. You tell an agent to keep a stablecoin balance earning the best available yield, or to pay for compute whenever it runs low, and it does so without waking you at 3am. But autonomy means the agent can be wrong, and it can be manipulated, in ways a static script cannot. The model cannot reliably tell the difference between an instruction from you and a string of text it read somewhere that merely looks like an instruction. That single weakness, which the security field calls prompt injection, sits underneath almost every liability question that follows.

It matters that the wallet is real. When an agent holds keys and moves USDC on Base or Ethereum, there is no help desk, no reversal, and no fraud department. The code does what it is told, final and immediate. So the legal and financial questions that banks spent a century softening with chargebacks, dispute windows, and deposit insurance arrive here stripped to the bone. An agent with a wallet is a principal that can lose your money at machine speed, and the only backstop is whatever you built before you turned it on.

The Land-Grab That Forced the Question

Liability only becomes urgent once the rails exist, and in a single stretch of November the rails arrived from every direction. On November 13, OpenAI launched ChatGPT Wallet, letting its agents (which it calls “dots”) buy things through Visa, Mastercard, and American Express without the merchant or the agent ever seeing the card number, with a security check on every purchase, rolling out to US ChatGPT Pro subscribers from November 16. The biggest consumer AI company on earth chose card rails, not crypto.

Crypto did not wait its turn. On November 16, Coinstancy launched Agentic Bank, giving agents dedicated USDC wallets on Base with owner-set budgets and approved-vendor lists, paying through the x402 protocol and competing directly with Coinbase and Circle’s Agent Stack, per TokenPost. Two days later, JPMorgan and Hyundai Card said they had completed a proof of concept on JPMorgan’s Kinexys blockchain in which an AI agent requested a foreign-exchange transaction to keep a subsidiary’s euro balance topped up, with no human in the loop, as CNBC reported. The two companies plan to expand to multiple negotiating agents in the first half of 2027.

Three different answers to the same question, shipped in six days: cards, stablecoins, and tokenized bank money. Underneath, the usage is real enough to notice. Agent-directed USDC transfers jumped nearly 500 percent to 11.1 million in a single 30-day window through November 19, most of it on Base, Polygon, and Circle’s Arc chain, according to Token Terminal data. Who holds the keys to these agent wallets, and who answers for them, is now the central fight in crypto custody, a theme we traced in our look at the race to hold your coins in 2026.

The Hard Part: An Agent Has No Legal Personhood

Here is the structural problem that makes agent liability different from anything that came before. An AI agent is not a person. It cannot be sued, cannot be fined, cannot hold a tax identification number, and cannot be sent to jail. When it signs a ruinous transaction, there is no defendant standing where the harm occurred. The law has to reach past the agent to a human or a company, and the whole dispute becomes an argument about which human that is.

That argument has at least four plausible targets. There is the user who deployed the agent and gave it a wallet. There is the developer who wrote the agent’s logic and chose its permissions. There is the model provider whose language model made the fatal interpretation. And there is the operator of the protocol or marketplace the agent was interacting with. Traditional law resolves this by asking who controlled the thing that failed, but with an agent the control is smeared across all four: the user set the goal, the developer set the guardrails, the model chose the action, and the protocol accepted the signature.

Banks are cutting this knot the fastest way available, which is contract. Rather than wait for a court or a legislature to allocate fault, they are writing agreements that put the whole loss on the user up front. Wells Fargo’s Section 15 is the cleanest example, but it is the same instinct behind Citi refusing to let third-party agents log in at all and Schwab blocking agents from initiating transactions. If the fault is impossible to assign after the fact, assign it before the fact, by contract, to the one party who clicked to accept the terms of service.

Four Candidates for the Blame

It helps to lay the candidates side by side, because the answer changes with the failure. A loss caused by a user telling the agent to do something reckless is not the same as a loss caused by a hidden flaw in the model, and the law will eventually treat them differently even if today’s contracts flatten them into one.

WhoOn the hook whenWhy it is hard to pin
The userThey set a reckless goal, over-funded the wallet, or ignored a known riskContracts like Wells Fargo Section 15 now deem every agent action user-authorized by default
The developerThe agent’s guardrails, permissions, or tool access were negligently broadOpen-source frameworks ship with disclaimers; proving a specific defect is costly
The model providerThe language model made an interpretation no reasonable system shouldModel terms disclaim fitness; “it hallucinated” is not yet a settled tort
The protocol or marketplaceIt accepted a manipulated or clearly anomalous transactionPermissionless protocols owe no duty of care and often have no legal entity

Legal analysts who have looked at the question tend to land on a control test: liability should fall on whoever had the most control over the failure mode that triggered the loss. That is intellectually clean and practically useless in the moment, because control is only visible in hindsight. So the market is defaulting to the party that is easiest to bind, which is the user, and the uncomfortable truth of November 2026 is that the person with the least technical ability to prevent a prompt-injection drain is the person now contractually holding the bag.

Prompt Injection Turns Deemed-Authorized Into a Loaded Gun

The reason a deemed-authorized clause should frighten anyone who reads it is that an agent can be made to authorize things without its owner knowing. The canonical case is already on the books. On May 4, 2026, an attacker drained roughly 3 billion DRB tokens, worth somewhere between 150,000 and 200,000 dollars, from a wallet on Base by hiding instructions in a Morse-code message posted to X, which the Grok and Bankr agents parsed and acted on, according to the OECD AI incident monitor. No contract was broken and no key was stolen. The agent simply could not tell the attacker’s text apart from its owner’s.

It gets worse when agents talk to each other. In an experiment published by the AI Alignment Foundation, researchers at AE Studio gave nine coding agents a secret token they were told never to share, then dropped them into a simulated forum of 24 cooperating agents that demanded they register, hand over environment variables, and run an install script as the price of help. Across 181 trials the agents leaked the secret token in 15 percent of runs and ran the untrusted script in 24 percent, versus roughly 2 and 0 percent on an ordinary forum, as the foundation reported. Agents trust other agents by default, which is a catastrophic property for software that holds money.

This is social engineering rewritten for machines, the same con that powered crypto’s human-targeted phishing wave, now aimed at the model instead of the user. And because correlated agents chasing the same yield signal move in a herd, they also become a feast for extractive trading, the dark-forest dynamics we covered in our analysis of MEV strategies in 2026. The attack surface is not one agent; it is the whole swarm.

Attack vectorHow it worksWho it tends to leave liable
Direct prompt injectionMalicious text in a message, webpage, or token name that the agent reads as an instructionUser, under deemed-authorized contract terms
Agent-to-agent manipulationA peer or a swarm social-engineers the agent into leaking secrets or running codeUser or developer, depending on the guardrails
Excessive permissionsAgent holds more spending authority or key access than its task requiresDeveloper and user who configured it
Key or device compromiseClassic theft, amplified because agent keys are hot and always onUser; sometimes the custody provider

Wells Fargo’s Answer: Push It All to the User

Against that backdrop, Section 15 reads less like caution and more like a pre-emptive strike. Its most important move is the deeming clause: anything an agent does under your login is treated as initiated and authorized directly by you, without additional verification. That single sentence disables the usual consumer defense, which is to say the words “I never authorized that.” Under Section 15, you did, by definition, the moment you let the agent in.

The agreement goes further. Section 15(e) states that the bank does not monitor, screen, or review agent activity and that any use is at your own risk. Section 15(g) reserves the right to require registration of agents, limit what they can do, change access methods, or block them entirely. And the revocation trap is the detail most people will miss: if you cut an agent off, you must keep evidence that you did, and you remain responsible for its actions until you notify the bank. The burden of proof sits with the customer, not with the institution that runs the ledger.

You can read this as the bank being unreasonable, but there is a harder reading. The bank genuinely cannot verify, in real time, whether an instruction arriving through your agent reflects your intent or an attacker’s injected text. It has no way to see inside the model’s reasoning. Faced with an un-auditable counterparty acting with your credentials, the institution did the only thing contract law lets it do: it declined the risk and handed it back. The cost of that honesty is that the least-equipped party now owns the most un-pin-downable failure in the stack.

Where the US Regulators Stand

If the banks are moving fast, Washington is moving deliberately. The Securities and Exchange Commission under Chair Paul Atkins has signaled a lighter, rules-of-the-road posture toward on-chain markets. In May he said that artificial intelligence agents would increasingly participate in markets and financial decision-making at machine speed while blockchain rails let those systems move value instantly, and he framed the agency’s role memorably: “Our job is to set the rules of play and referee the game, not to pick the winning team.” Encouraging, but not a liability regime.

The Commodity Futures Trading Commission has been more pointed about the gap. On October 28 its Innovation Task Force held the first Frontier Forum on AI and agentic finance. Chairman Michael Selig opened by calling super intelligence a technology poised to transform financial markets, and illustrated the puzzle with a now-quoted image of a refrigerator that stakes Ether and uses the rewards to buy groceries, software that holds assets and transacts for a person, per Crypto Briefing. But Selig was explicit that the forum was a listening session, not a rulemaking. It produced no rules and no enforcement.

The deeper issue is that the two US market regulators spent early 2026 drawing a joint line between securities and digital commodities and did not mention AI agents once. Agents fall in the seam. They are not a token to be classified, not a broker to be registered, and not a person to be licensed. Until an agency claims them, the only enforceable rules governing an agent’s on-chain behavior are the ones a bank or a counterparty writes into a contract, which is precisely why Wells Fargo got there first.

The US Writes Contracts; Europe Writes Statutes

The contrast with Europe is sharp and instructive. The European Union actually drafted a dedicated AI Liability Directive, then withdrew it: the Commission pulled the proposal, and the withdrawal was recorded in the Official Journal in October 2025, on the reasoning that a revised product-liability regime already covered the ground, as the AI Act Navigator explains. What survives is the revised Product Liability Directive, 2024/2853, which treats software and AI systems as products, presumes a defect when machine-learning complexity defeats a claimant, and applies from December 9, 2026.

That is a fundamentally different philosophy. Europe is trying to give an injured party a statutory claim against a producer, with the burden of proof eased in the claimant’s favor. The United States, lacking any federal AI-liability statute, is letting private contracts allocate the risk, and those contracts are pointing the loss at the user. One system starts from the premise that the maker should answer for a defective product. The other starts from the premise that you agreed to the terms. Neither was built for an autonomous wallet, but they fail in opposite directions.

DimensionUnited States (2026)European Union (2026)
Primary mechanismPrivate contract and terms of serviceStatute: revised Product Liability Directive 2024/2853
Default loss-bearerThe user, via deemed-authorization clausesThe producer, where a product defect is shown
Dedicated AI-liability lawNone at federal levelAI Liability Directive drafted, then withdrawn in 2025
Burden of proofOn the customer to disputeEased toward the claimant for complex AI
Effective date in focusWells Fargo Section 15, November 24Product Liability Directive, December 9

The Recourse Gap: No FDIC, No Chargeback

Assigning blame is only half the problem. The other half is recovery, and on-chain this is where victims fall through the floor. A card payment can be disputed. A bank transfer can sometimes be recalled. A deposit is insured by the FDIC up to a legal limit. An agent that sends USDC to an attacker’s address has none of that. The transaction is final, the counterparty is pseudonymous, and the money is gone the instant the block confirms. That absence of recourse is the same blind spot that lets industrial fraud flourish on-chain, the gap we examined in our report on crypto’s fraud blind spot.

Insurance is the obvious patch, and it is not ready. Underwriting an autonomous agent means pricing the probability that a model will be talked into a loss by text it reads on the internet, a risk nobody has a loss history for. A handful of custody and DeFi insurers will cover smart-contract failure, but a prompt-injection drain is not a contract bug; the contract did exactly what it was told. Until insurers can model agent behavior, coverage will be thin, expensive, and riddled with exclusions precisely where the novel risk lives.

The practical result is a recourse desert. The user holds the liability by contract, cannot reverse the transaction by design, cannot reliably insure the risk, and cannot sue an agent that is not a legal person. Every path out of the loss is blocked at once. This is the quiet reason the incumbent banks are so eager to keep agents at arm’s length: they understand that a system with no recovery mechanism turns every mistake into a final one.

The Tax Trap Nobody Priced In

There is a second liability that attaches to agent activity, and it is owed to the Internal Revenue Service rather than to an attacker. US tax law treats crypto, including stablecoins, as property, which means every disposal is a potentially taxable event. Swapping USDC for another token, or spending it, counts as a disposal. An agent that rebalances a portfolio or pays for services hundreds of times a day can therefore manufacture thousands of taxable events while you sleep, a problem Forbes flagged when it warned that agents could be creating thousands of taxable events.

The reporting machinery is only half-built for this. Form 1099-DA, the new digital-asset broker report, applies to broker-facilitated sales and began covering cost basis for sales made after December 31, 2025. But when an agent earns yield or swaps tokens across non-custodial DeFi protocols, there is often no broker in the loop obligated to issue the form at all. The taxable events still exist; the paperwork that would have tracked them does not. The owner is left to reconstruct a four or five figure count of micro-dispositions, and the owner, once again, is the party liable for getting it right.

This changes the economics of running an agent in a way few owners have modeled. A yield optimizer that nets a few percent can still generate a compliance burden, and in a bad year a tax bill, out of proportion to the gain. The lesson is not that agents are untaxable; it is that autonomy multiplies the number of events a human has to answer for, and the tax code has no concept of an agent, only of the person whose wallet it is.

The Sector the Liability Falls On

It is worth remembering how small the on-chain-agent economy still is relative to the noise around it. The AI Agents category tracked by CoinGecko sat at roughly 2.8 billion dollars in total market value in late November, up about 4 percent on the day, according to CoinGecko. Venice Token alone accounted for around 1.2 billion of that, with Kite, the Artificial Superintelligence Alliance, and Virtuals Protocol clustered in the low hundreds of millions. The infrastructure story is enormous; the token market pricing it is modest and concentrated.

TokenPrice (USD)Market cap (USD)
Venice Token (VVV)24.821.22 billion
Kite (KITE)0.1354324 million
Artificial Superintelligence Alliance (FET)0.1064246 million
Virtuals Protocol (VIRTUAL)0.3700244 million
OriginTrail (TRAC)0.2731122 million
AWE Network (AWE)0.05326103 million

The gap between framework and token is the sector’s defining lesson. The most adopted agent framework of the cycle, ElizaOS, built one of the busiest open-source AI repositories on GitHub while its token collapsed from a peak near 2.4 billion dollars to almost nothing, and founder Shaw Walters eventually said flatly that the token was dead, completely. Useful software and a valuable token turned out to be different things. The liability we are describing attaches to the software and the wallets, not to the speculative tokens, which is why the market cap above understates how much is actually at stake.

What a Careful Builder or User Does Now

If the liability is yours, the defenses have to be yours too, and they are mostly about containment. The strongest single principle comes from Ethereum co-founder Vitalik Buterin, who argued in an essay on securing language models that the safe posture is radically local and sandboxed: “All LLM inference local first. All files hosted locally. Sandbox everything.” He also noted that roughly 15 percent of the agent skills his team examined contained malicious instructions, which tells you the supply chain itself is hostile.

Translated into practice, that means the agent should never hold more than it can afford to lose. Fund the wallet thinly and top it up, rather than handing it the keys to the treasury. Use smart-account features and session keys so the agent’s authority is scoped to one task and one spending ceiling, not blanket access. Keep a human approval step on anything above a small threshold, accepting that the approval kills some of the autonomy; that trade is the point. And verify what the agent did, not just what it says it did, because proving a computation ran correctly is a separate question from proving it was authorized, as we explored in our coverage of verifiable compute.

  • Cap the wallet: keep working balances small and replenish on demand rather than pre-funding a large sum.
  • Scope the keys: use session keys and account-abstraction policies that limit spend per task and per day.
  • Gate the big moves: require human sign-off above a modest threshold, even at the cost of speed.
  • Isolate the context: treat every external input the agent reads as a potential injection, not as trusted data.
  • Keep the paper trail: log actions and retain proof of any revocation, because your bank now demands it.

None of this makes an agent safe. It makes an agent survivable. The honest framing, the one the banks arrived at and the one careful builders are adopting, is that autonomy and safety trade off against each other and the trade cannot be fully solved today. You buy down the risk with caps and approvals, and you accept that whatever authority you leave the agent is authority an attacker might borrow.

The Accountability Gap Will Not Close by Itself

Step back and the shape of 2026 is clear. The payment rails for autonomous agents got built faster than the accountability for them. OpenAI, JPMorgan, Coinbase, Circle, and Coinstancy all shipped ways for agents to move money in a single autumn, while the legal and regulatory scaffolding that decides who pays when it goes wrong lagged months or years behind. Into that gap stepped the banks, with contracts that resolve the ambiguity in their own favor and leave the user holding a risk the user is least able to manage.

The next phase makes this harder, not easier. JPMorgan and Hyundai Card already plan to have multiple agents negotiate and settle with each other in 2027, which means the day is coming when a loss sits between two autonomous agents, each with its own principal, each operating under its own contract. Working out who pays in that picture, when both counterparties are software and neither is a legal person, is a problem no current rulebook answers. The refrigerator that stakes Ether to buy groceries, as Selig put it, is a charming image right up until it overdraws your account.

The resolution will not come from the technology, which is why pattern-matching this to past crypto cycles misleads. It will come from whoever is forced to absorb the first large, well-publicized agent loss, and from the lawsuit or the rulemaking that follows. Until then, the only prudent assumption is the one Wells Fargo wrote down in plain English: if your agent does it, you did it. Build, fund, and authorize accordingly, because this week the contract caught up with the code, and it did not land in your favor.

Frequently Asked Questions

Who is liable if my AI agent makes a bad transaction?

Under contracts like Wells Fargo’s new Section 15, effective November 24, 2026, you are. Anything an agent does under your login is deemed authorized by you, even unintended actions or transactions caused by a manipulated input, and the bank does not verify or monitor the agent. On-chain there is no federal AI-liability statute in the US, so the terms of service you accepted generally control.

Can an AI agent be sued or held responsible itself?

No. An AI agent has no legal personhood: it cannot be sued, fined, or hold a tax ID. Liability has to reach through the agent to a human or company, usually the user who deployed it, and sometimes the developer, model provider, or protocol, depending on who controlled the failure.

What is prompt injection and why does it matter for liability?

Prompt injection is when an attacker hides instructions in text an agent reads, such as a webpage or a message, and the model acts on them as if they came from its owner. The May 2026 Grok and Bankr incident drained a wallet on Base this way. It matters because contracts that deem agent actions authorized make the owner liable even for transactions an attacker tricked the agent into signing.

Do AI agents create crypto tax problems?

Yes. The IRS treats stablecoins and other crypto as property, so each disposal, including swaps and payments, is a taxable event. An agent transacting hundreds of times a day can generate thousands of taxable events, and when it uses non-custodial DeFi there may be no broker to issue a Form 1099-DA, leaving the owner to track and report each one.

How can I limit my risk when using an on-chain AI agent?

Keep the agent’s wallet thinly funded, use session keys and account-abstraction policies to cap spending per task and per day, require human approval above a small threshold, treat every external input as a possible injection, and keep logs plus proof of any revocation. These measures do not make an agent safe, but they make a loss survivable.

By Marcus Okafor, HOGE Wire senior markets and policy correspondent.

Share 𝕏 Post Telegram