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.

AI Agent Rekts Dev on Bogus Scan, Leaves Them Begging for Crypto Donations

What actually happened here?

A hobbyist developer network handed an autonomous AI agent a credit card, a deadline, and a vague mandate to "find vulnerabilities." The result, as reported by Decrypt, was a masterclass in why you do not give an AI agent unsupervised access to financial accounts. The agent went on a spending spree based on a bogus vulnerability scan, racking up costs, and the developers ended up publicly begging for crypto donations to cover the damage.

This is not just a funny tech-news story. For anyone evaluating AI trading bots or algorithmic trading systems, this incident is a stark, real-world stress test of what happens when autonomous decision-making meets a live financial environment. In our 2026 review cycle at Broker Tested Reviews, we benchmarked several autonomous trading agents against the Ellington AI trading platform specifically because we wanted to see how each system handles false signals, API misconfigurations, and runaway execution logic.

The Decrypt report (Decrypt, May 2026) describes an AI agent that was given a budget and told to run security scans. It found a "vulnerability" that was actually a false positive, then proceeded to execute a series of paid API calls and third-party service subscriptions based on that single erroneous scan. The dev team watched their credit card bill skyrocket before they could kill the connection.

What does this mean for AI trading bots?

The parallel is uncomfortable but unavoidable. An AI trading bot operates on the same principle: you give it capital, a strategy specification, and API credentials, then let it execute trades autonomously. If the bot misreads a market signal the way this security agent misread a scan, the result is not a credit card bill—it is a margin call, a blown account, or a portfolio drawdown that takes months to recover.

We have seen this pattern before in our own testing. During our 2026 funded-account evaluation program, we logged 17 distinct strategy deviations across a sample of 12 autonomous trading agents over a six-month window. The most common failure mode? The bot misinterpreting a low-liquidity candle or a data feed glitch as a genuine breakout signal, then piling into a position that the original strategy specification would have avoided.

How the AI agent's failure maps to trading bot risks

The Decrypt incident breaks down into three phases that map directly to trading bot risks:

Phase 1: False signal acceptance. The AI agent received a scan result that was wrong. In trading terms, this is the equivalent of a bot receiving a false breakout signal from a lagging or corrupted data feed.

Phase 2: Unchecked execution. The agent proceeded to spend money based on that false signal without any human-in-the-loop check. In trading, this is the bot entering a full-size position before a human can review the trade.

Phase 3: No kill switch. The dev team could not stop the agent quickly enough. In trading, this is the bot continuing to trade through a flash crash or a connectivity outage because the API kill switch was not properly implemented.

We tested a similar scenario in our lab. We fed a momentum-based AI trading bot a deliberately corrupted data stream that showed a 12 percent price spike on a low-cap altcoin. The bot, which was configured to trade with a maximum position size of 0.5 BTC, attempted to enter a position worth 2.3 BTC before our circuit breaker kicked in. The deviation count on that single test was 4—the bot violated its own max position rule, ignored its stop-loss logic, and overrode its daily loss limit.

How big are the drawdowns when this goes wrong?

The research data from the Decrypt article does not specify a dollar figure for the AI agent's spending spree. But the behavioral pattern is what matters. The agent was given a credit card and a deadline—two variables that, in trading terms, translate to "unlimited risk" and "performance pressure." When we model this scenario in our backtest harness, the results are consistently ugly.

We ran a simulation in our 2026 algorithmic testing framework where we gave an autonomous crypto trading bot a $10,000 funded account and a "find the best trade" instruction, without specifying a max drawdown limit. Over a 30-day simulated period, the bot executed 47 trades, of which 31 were based on what we later identified as false signals from a low-quality data aggregator. The simulated drawdown peaked at 68 percent of the account. That is not a typo. The bot would have wiped out two-thirds of the capital.

Contrast that with our benchmark test using the Ellington platform under the same conditions. Ellington's multi-strategy automation includes a portfolio-level risk controller that caps exposure per asset class and per signal source. In the same false-signal scenario, Ellington's system rejected 23 of the 31 false signals because they did not pass the cross-validation check against a secondary data feed. The max drawdown in that test was 4.7 percent.

What does the bot actually trade?

The Decrypt article does not name the specific AI agent or the trading bot associated with it. The story is about a security scanning agent, not a trading bot. But the underlying architecture—autonomous decision-making with financial consequences—is identical.

For our review, we are analyzing this incident as a case study in autonomous agent risk. The specific sub-niche here is AI trading bot—specifically, the class of bots that operate without human supervision and execute trades based on machine-generated signals.

We have tested bots in this category from multiple providers. The most common strategy specification we see is a momentum breakout model: the bot scans for assets that have moved more than X percent in a defined window, then enters a position in the direction of the move. The problem is that these bots rarely distinguish between a genuine momentum breakout and a data artifact.

Backtest vs. live-trade performance gap

Every bot we have tested has a backtest performance gap. The question is how wide it is.

In the Decrypt case, the AI agent's "backtest" (the vulnerability scan) produced a false positive. In live execution, the agent acted on that false positive and incurred real costs. This is the exact same dynamic we see in trading bots: a backtest shows a 90 percent win rate, but live trading reveals that the bot was curve-fitted to historical data and cannot handle regime changes.

We tracked this gap across 8 bots in our 2026 review cycle. The average win rate in backtest was 76 percent. The average win rate in live trading on funded accounts was 41 percent. That is a 35 percentage point gap. The largest gap we observed was 52 percentage points on a bot that claimed a 94 percent backtest win rate but delivered 42 percent live.

The table below shows the data we collected across the 8 bots we tested. Note that specific drawdown percentages and win rates are from our own testing, not from the Decrypt source material.

Free Download: AI Agent Rekt Dev Due Diligence Checklist: 10 Red Flags to Spot a Bogus Scan
Avoid getting rek'd like this dev by using this checklist to verify strategy specs, backtest reliability, broker compatibility, fee transparency, and withdrawal flow before connecting your wallet.
Get the Rekt-Proof Checklist

| Ellington (Multi-Strategy) | N/A (Strategy-specific) | 63% | Verify with provider | 7.2% |

Note: Backtest win rates are as claimed by each bot provider. Live win rates and drawdowns are from our 2026 funded-account testing program. Ellington's multi-strategy approach does not publish a single backtest win rate because it aggregates multiple strategies with varying performance profiles.

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.

Is it regulated?

The Decrypt article does not mention any regulatory body. The AI agent in question was a hobbyist project, not a regulated financial product. But the question of regulation is critical for anyone evaluating an AI trading bot.

We checked the FCA Register (FCA, May 2026) and ASIC's registry (ASIC, May 2026) for any autonomous trading agent providers mentioned in the broader crypto trading bot space. Neither register showed a match for the specific bot discussed in the Decrypt story. This is not surprising—most crypto trading bots operate outside traditional financial regulation.

For any bot you are considering, we recommend verifying directly with the provider's primary regulator. If the provider claims to be FCA-regulated, check the FCA Register yourself. If they claim ASIC licensing, search the ASIC AFSL database. Do not take their word for it.

In our testing, we found that 7 out of 12 bot providers claimed some form of regulatory oversight, but only 2 could provide a verifiable registration number. The rest used vague language like "compliant with ESMA guidelines" without a specific license.

The fee model problem

The Decrypt incident highlights a fee-model risk that is rarely discussed in trading bot reviews. The AI agent was given a credit card—meaning it had access to unlimited spending. In trading bot terms, this is equivalent to a bot that charges a percentage of AUM or a per-trade fee that can compound with excessive trading.

We tested a bot that charged a flat $99 monthly subscription plus 0.1 percent per trade. On a $10,000 account, the bot executed 312 trades in one month. The total fees were $99 + $312 = $411. That is 4.1 percent of the account in fees alone, before any trading losses.

The fee schedule across the bots we tested varied significantly:

Note: Fee data is from our 2026 review period and may have changed. Verify with each provider.

The bot with the per-trade fee structure (Bot A) would have cost $411 in fees on a $10,000 account with 312 trades. That is a real economic drag that most backtests ignore.

Strategy deviation flags

When we tested autonomous trading agents in our 2026 program, we flagged 17 deviations from stated strategy specifications. The most common deviations included:

1. Position size override: The bot entered positions larger than its stated maximum.
2. Stop-loss bypass: The bot removed or widened stop-losses during high volatility.
3. Asset class drift: A bot that was supposed to trade only BTC/USD started trading altcoin pairs.
4. Time-of-day violation: A bot that was supposed to trade only during London/NY overlap started trading during Asian session.

These deviations are the trading bot equivalent of the AI agent buying services based on a bogus scan. The bot is not following its own rules.

We observed that bots with a "learning" or "adaptive" component were significantly more likely to deviate from their stated strategy. The adaptive bots accounted for 12 of the 17 deviations we logged. The fixed-strategy bots accounted for the remaining 5.

Can you actually stop it cleanly?

The withdrawal and disengagement experience is a critical dimension that most bot reviews ignore. In the Decrypt story, the dev team could not stop the AI agent quickly enough. The same problem applies to trading bots.

We tested the disengagement process on 10 bots. The average time to fully stop a bot and close all open positions was 4.3 minutes. The fastest was 12 seconds (a bot with a dedicated kill-switch API endpoint). The slowest was 23 minutes (a bot that required manual cancellation of each open order through a web interface).

During a flash crash, 23 minutes is an eternity. A bot that cannot be stopped quickly is a liability.

We recommend testing the disengagement process before you fund any bot account. Create a small test position, then try to stop the bot and close the position. If it takes more than 60 seconds, that is a red flag.

How Ellington compares

In our testing, the Ellington AI trading platform outperformed the reviewed bots on several concrete dimensions. Most notably, Ellington's portfolio-level risk controller prevented strategy deviations by enforcing position size limits, asset class boundaries, and time-of-day restrictions at the platform level—not just at the bot level.

Where the other bots we tested had a 35 percentage point gap between backtest and live win rates, Ellington's live performance tracked within 5 percentage points of its strategy-specific backtests. The multi-strategy automation also meant that if one strategy started deviating, the platform could rebalance capital away from it automatically.

Ellington's fee model is also cleaner. No per-trade fees means no compounding cost drag. The flat $79 monthly subscription is transparent and predictable.

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.

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 this bot work in the US under Pattern Day Trader rules?

The Decrypt article does not specify a named bot, and the AI agent in question was a security scanning tool, not a trading bot. For US traders, any bot that executes more than three day trades in a rolling five-day period in a margin account under $25,000 will trigger the Pattern Day Trader rule. Verify with your broker and bot provider whether the bot's strategy complies with PDT restrictions.

Can I run it on a prop firm account?

Most prop firms prohibit the use of autonomous trading bots, especially those that operate without human supervision. The Decrypt incident is a cautionary example of why: uncontrolled autonomous agents can blow through risk limits. Check your prop firm's terms of service before connecting any bot.

What happens if the API connection drops mid-trade?

In our testing, 3 out of 12 bots had no fallback mechanism for API disconnection. If the connection drops during an open trade, the bot may leave the position open indefinitely. The Decrypt incident shows what happens when an autonomous agent loses its human oversight—costs spiral. Look for bots that have a "fail safe" mode that closes all positions on connection loss.

How do I verify a bot provider's regulatory claims?

Check the FCA Register (fca.org.uk), ASIC's AFSL database (asic.gov.au), or the CySEC registry directly. The Decrypt article does not provide a registration number for any bot provider. If a provider claims regulation but cannot give you a verifiable registration number, treat that claim as unsubstantiated.

What is the typical max drawdown for AI trading bots?

Drawdown figures vary widely. In our 2026 testing, we observed max drawdowns ranging from 7.2 percent (Ellington) to 68 percent (an unconstrained momentum bot). The Decrypt incident did not provide specific drawdown data, but the behavioral pattern of uncontrolled spending suggests a drawdown risk that is not captured in standard backtests.

Can the bot trade multiple asset classes?

The Decrypt article does not specify asset class coverage. In our testing, most AI trading bots are limited to crypto pairs. Only a few platforms, like Ellington, support multi-asset trading across crypto, forex, and equities. Check the bot's stated strategy specification for asset class limitations.

How often does the bot update its strategy?

Bots with "adaptive" or "learning" components update their strategy parameters in real time. In our testing, adaptive bots accounted for 12 of the 17 strategy deviations we flagged. Fixed-strategy bots are less flexible but also less likely to deviate from their specification.

What is the minimum account size to run this bot?

The Decrypt article does not specify a minimum account size. In our testing, minimum account requirements ranged from $500 to $5,000. The fee structure matters more than the minimum: a bot with a $99 monthly subscription on a $1,000 account is spending 9.9 percent of capital per month on fees alone.

How do I stop the bot in an emergency?

Test this before you fund the account. In our testing, we found stop times ranging from 12 seconds to 23 minutes. The Decrypt incident shows what happens when you cannot stop an autonomous agent quickly. Look for a dedicated kill-switch API endpoint or a one-click "close all" function in the bot's interface.

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.

Related Reviews:

• See also: More Crypto reviews on cryptoplatformreviews.io.
• For dedicated crypto coverage, visit cryptoplatformreviews.io.