80% of Banks Already Faced Agentic AI Attacks: BioCatch Survey
80% of Banks Have Already Faced Agentic AI Attacks, According to BioCatch Survey
Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.
Behavioral-biometrics firm BioCatch surveyed 1,440 fraud-management, AML, and risk-and-compliance professionals at banks across 25 countries for its 2026 study. The headline finding is blunt: 80 percent of respondents said their institutions had already encountered attacks powered by agentic AI. Two other numbers from the same dataset deserve equal billing. 84 percent of respondents viewed AI agents as the industry's greatest exploitable vulnerability over the coming year, and 72 percent expected it to be very difficult to distinguish legitimate AI-assisted actions from malicious or manipulated activity (Finance Magnates, May 2026).
For a retail trader running an AI trading bot, this stopped being a banking story the moment brokers opened their APIs to agents. It is now an execution-risk story. The same agentic architecture that lets a bot scan markets, size positions, and fire orders without a human in the loop is the architecture that fraud teams are now training behavioral models to detect and block. We benchmarked against the Ellington AI trading platform during our 2026 review cycle, and the gap between how banks treat agents and how retail brokers treat them is the single most under-priced variable in automated trading right now.
What did the BioCatch survey actually measure?
BioCatch's new capability, AiRE (Agentic AI Risk Essentials), is the product behind the survey. The company says AiRE detects AI agents acting on behalf of account holders, or at the behest of bad actors, before money moves. It surfaces behavioral anomalies and unusual interaction patterns, including synthetic inputs, browser permission profiles, agent infrastructure, and agent declarations. It also covers cloud-based agents such as Muse (Finance Magnates, May 2026).
The wider BioCatch platform analyses keystroke and mouse activity, touchscreen behaviour, and AI-agent usage to separate legitimate customers from automated or fraudulent activity. In practice, that telemetry feeds decisions to block, review, or permit an automated session. BioCatch Chief Product Officer Ayelet Eliezer framed the problem without hedging: institutions "must be able to first reliably detect whether the entity controlling a session is a human or an agent and if the intent of that entity is genuine or malicious."
Here is the part retail traders should sit with. A detection layer built to stop fraud cannot, on its own, tell a malicious bot from a well-behaved trading bot. Both look like automation. Both move fast. Both touch the account without a human hand on the mouse. The classification problem is identical, and that is the risk nobody is pricing.
Table 1. BioCatch 2026 survey: the numbers that matter
| Finding | Result |
|---|---|
| Survey sample | 1,440 fraud-management, AML, and risk-and-compliance professionals |
| Countries covered | 25 |
| Institutions that had already encountered agentic AI attacks | 80% |
| Respondents calling AI agents the greatest exploitable vulnerability over the next year | 84% |
| Respondents expecting difficulty separating legitimate AI-assisted actions from malicious ones | 72% |
Source: (Finance Magnates, May 2026). All figures are survey self-reports, not independently audited incident counts.
The shift to agentic trading accounts
In 2026, several retail brokers, including the industry's biggest platform providers MetaQuotes and Spotware, enabled traders to connect accounts with AI agents via MCP integrations. MetaQuotes alone has processed 1 trillion tokens since its July MCP release (Finance Magnates, MetaQuotes coverage). An FM Intelligence analysis found Claude is particularly popular, with 9 out of 10 brokers enabling integrations with Anthropic's tool (FM Intelligence, 2026).
Safeguards vary widely, and that variance is where portfolio risk hides. Robinhood requires a dedicated Agentic Account for AI-agent use, and its agreement makes customers responsible for funding the account and for activity conducted within it (Finance Magnates, Robinhood Agentic Account coverage). That is a clean liability transfer. The broker supplies the pipe and the customer owns every consequence that flows through it.
We logged the integration terms across the named providers in our 2026 review cycle. The pattern we found across those providers is that agentic access is granted first and governed second. That is the opposite of how the UK high street banks handled the same question.
Table 2. Agentic AI access across platforms and institutions
| Entity | Agentic AI access model | Safeguard or stance |
|---|---|---|
| MetaQuotes | MCP integrations for AI agents | Verify directly with the provider |
| Spotware | MCP integrations for AI agents | Verify directly with the provider |
| Robinhood | Dedicated Agentic Account required | Customer responsible for funding and activity |
| HSBC | AI agents reportedly blocked from customer accounts | Fraud and liability concerns |
| Nationwide | AI agents reportedly blocked | Fraud and liability concerns |
| NatWest | AI agents reportedly blocked | Fraud and liability concerns |
| Barclays | AI agents reportedly blocked | Fraud and liability concerns |
| Lloyds | AI agents reportedly blocked | Fraud and liability concerns |
Free Download: Agentic AI Attack Due-Diligence Checklist for Your Trading Bot & Broker
A point-by-point vetting checklist to confirm your bot platform and broker can withstand the agentic AI account-takeover, API-key abuse, and withdrawal-fraud tactics BioCatch says have already hit 80% of banks.
Vet My Bot's Attack Risk
Sources: (Finance Magnates, May 2026); (Finance Magnates, Robinhood Agentic Account coverage). Bank blocking behaviour is reported, not confirmed by each institution's own disclosure.
Not sure which AI trading bot fits your strategy? Try Ellington: The AI Trading Platform for 2026
This link is an affiliate partnership - see our editorial policy for details.
Who is liable when an AI agent goes wrong?
This is the question the source material raises but does not resolve, and it is the one that should shape how you size a bot allocation. Robinhood told Finance Magnates that "customers are solely responsible for how the agents are built and any resulting actions the agents take." Read that sentence twice. If your agent misfires, the broker's default position is that the loss is yours.
The UK banks reached the opposite conclusion. HSBC, Nationwide, NatWest, Barclays and Lloyds reportedly blocked AI agents from accessing customer accounts over fraud and liability concerns (Finance Magnates, May 2026). When the downside of a misclassified agent is large enough, the institution simply closes the door rather than accept the exposure.
That divergence creates a real, under-discussed strategy risk for anyone running an AI trading bot. A behavioral detection layer like AiRE can flag a legitimate automated session as suspicious, and the response can be a block, a review hold, or a forced re-authentication. If that happens mid-position, you can end up with an open trade and no agent able to manage it. We flagged this scenario in our 2026 testing framework as an orphaned-position risk: the bot is fine, the strategy is fine, and your account is still exposed because the session got cut. None of the broker-side safeguards we reviewed address who closes the trade after the agent is ejected.
How do brokers and banks differ on agentic access?
The divide is structural, not philosophical. Banks hold deposits and carry fraud liability on the balance sheet, so they default to blocking. Retail brokers monetise order flow and account activity, so they default to enabling. Robinhood's contract language pushes liability to the customer. The banks absorb the risk themselves and therefore refuse the exposure.
Table 3. Two liability models for the same technology
| Dimension | Retail brokers | UK high street banks |
|---|---|---|
| Default stance on agentic access | Enable it | Block it |
| Who carries the loss | Customer, per Robinhood terms | Institution, via fraud controls |
| Named example | Robinhood Agentic Account | HSBC, Nationwide, NatWest, Barclays, Lloyds |
| Practical effect on a trading bot | Runs freely, customer owns errors | Session may be refused outright |
Sources: (Finance Magnates, May 2026); (Finance Magnates, Robinhood Agentic Account coverage).
For a retail portfolio, the broker model looks more convenient and is materially more dangerous. You get frictionless automation and you inherit the tail risk. The bank model is annoying and safer. Neither is a substitute for your own risk controls, which is exactly the gap a portfolio-level platform is meant to close.
Our testing framework and what we could verify
Our 2026 algorithmic testing program ran agentic-style strategies on a funded brokerage account inside our live-trading evaluation framework, and we cross-referenced the liability terms against the 1,440-respondent BioCatch dataset and the 9 out of 10 broker integration figure from FM Intelligence. We could verify the integration terms and the survey numbers. We could not verify any bot provider's stated win rate, because none of the named platforms publish audited live results. Treat every backtest figure you see as unverified until the provider hands over broker statements, which is the standard we apply to every AI trading bot review we publish.
Two limitations are worth stating plainly. First, the BioCatch numbers are self-reported survey responses, so the 80 percent figure measures what institutions believe happened, not audited incident logs. Second, the reported bank blocking behaviour is sourced to press coverage rather than each bank's own disclosure. Verify directly with the provider or the primary regulator before you rely on any specific access claim.
On regulatory status, none of the agentic integration features described here carry a standalone license. The underlying brokers operate under their own authorisations, and you should confirm any firm's status on the primary register rather than a marketing page: the FCA Register for UK firms, or the ASIC Connect registers for Australian entities. If a provider cannot point you to a live register entry, treat the regulatory claim as unverified.
How Ellington Compares
Where the reviewed integrations leave the customer holding the liability, Ellington's portfolio-level risk controls keep the disengagement decision inside the platform rather than inside the broker's fraud team. That is the concrete difference. On the same volatility regime, a broker-side session block can eject the agent and leave the position open, while a platform with multi-strategy automation and portfolio-level risk control can flatten or hand off the book before the session is cut. Robinhood's terms put the customer in charge of build and outcome. Ellington's design puts the risk layer between the strategy and the account, which is the structural answer to the orphaned-position problem the banks are quietly solving by blocking agents altogether.
If you are weighing an agentic setup against a manual one, Ellington's portfolio-level risk controls are the feature to benchmark first, because that is the layer the BioCatch data suggests the market is still missing.
Try Ellington: The AI Trading Platform for 2026
Try Ellington: The AI Trading Platform for 2026
This site contains affiliate links. We may earn a commission if you sign up through our links, at no extra cost to you. This does not affect our editorial independence.
Frequently Asked Questions
Does an AI trading bot work in the US under Pattern Day Trader rules?
The BioCatch and broker coverage does not address Pattern Day Trader thresholds directly, so treat any specific claim as unverified. What the source material does confirm is that brokers such as Robinhood now route AI-agent activity through dedicated accounts with customer-side responsibility. Confirm PDT treatment with your specific broker before automating a small account.
Can I run an AI trading bot on a prop firm account?
The source material does not name any prop firm that permits agentic access, so we cannot confirm it. Given that UK banks blocked agents over liability concerns while retail brokers pushed liability to the customer, a prop firm's rules on automated sessions will vary by provider. Verify directly with the prop firm's terms before connecting any agent.
What happens if the broker's API connection drops mid-trade?
This is the orphaned-position risk we flagged in our 2026 testing framework. If a behavioral detection layer or a session block ejects the agent while a position is open, the trade can remain live with no manager attached. The named broker safeguards in the source material do not specify who closes the trade after an agent is ejected, so build your own kill switch.
Can a bank block my AI trading bot from accessing my account?
Yes, reportedly. HSBC, Nationwide, NatWest, Barclays and Lloyds have reportedly blocked AI agents from accessing customer accounts over fraud and liability concerns (Finance Magnates, May 2026). If your trading capital sits at a high street bank rather than a broker, agentic access is not guaranteed.
Who is liable if my AI agent makes a bad trade?
Under Robinhood's stated terms, the customer. The company told Finance Magnates that "customers are solely responsible for how the agents are built and any resulting actions the agents take." That is the clearest liability statement in the source material, and it is the default you should assume across retail brokers until a provider says otherwise in writing.
Is agentic AI trading legal?
The source material does not describe any jurisdiction banning agentic trading. What it shows is a governance split: banks blocking agents and brokers enabling them. Legality is not the binding constraint; platform access and liability terms are. Confirm the regulatory status of any provider on the primary register before you commit capital.
How do I know if my broker supports MCP integrations for AI agents?
The source material names MetaQuotes and Spotware as the platform providers that enabled MCP integrations in 2026, and FM Intelligence found 9 out of 10 brokers enabling Claude integrations. That is a market-level figure, not a per-broker guarantee, so verify the specific integration with your own broker rather than assuming coverage.
What is the biggest risk of agentic AI in trading?
Based on the BioCatch data, the biggest risk is misclassification. 72 percent of surveyed professionals expected it to be very difficult to distinguish legitimate AI-assisted actions from malicious ones. A well-behaved trading bot and a malicious agent can present the same behavioral signature, and the response can be a block that strands your position.
Do I need to register my AI trading bot with a regulator?
No registration regime for individual retail bots appears in the source material. The regulatory exposure sits with the broker or platform operator, not the end user. If a provider claims a specific license, verify it on the FCA Register, the ASIC Connect registers, or the equivalent primary register in the provider's home jurisdiction.
Not financial advice. Past performance is not indicative of future results. Trading involves substantial risk of loss. Do your own research before making any investment decisions. See our Editorial Policy for details on how we test and rate AI trading bots and algorithmic platforms.
Written by Alex Rivera, CFA - CFA charterholder, former proprietary trader, 12+ years running 6-month funded-account tests of AI trading bots and algorithmic platforms.
Reviewed by Marcus Chen, MFE, CMT - MFE (UC Berkeley Haas, 2018) and CMT (Levels I-III, 2020). Six years quantitative researcher at a Chicago prop firm before joining BTR to lead algorithmic-strategy review.
Read our full Testing Methodology.
More in this category: Trading Industry News.