An AI risk firewall is a piece of infrastructure that sits between a copy-trading publisher and your wallet. Every signal the publisher sends gets scored against an independent risk model before it executes on your account. Bad signals get blocked. Good ones pass through.
This sounds like a small thing. In practice it's the difference between modern copy trading and the 2010-era version, where the publisher's worst day was also your worst day.
This post walks through how Foxy AI β BottomUP's risk firewall β actually works: what data it reads, how the score is built, and what gets blocked vs. what passes through.
Why the firewall exists
The classic copy-trading failure mode is the revenge trade. A publisher takes a loss, gets emotional, doubles their position size on the next entry. If you copy them at 1Γ, your account takes the doubled loss too. The mirror is faithful β that's the problem.
Three failure modes show up over and over in post-mortems:
- Revenge trades β emotional response to a recent loss
- Leverage drift β trader quietly moves from 3Γ to 20Γ
- Strategy abandonment β momentum trader silently switches to mean-reversion mid-month
A 2010-era platform mirrors all three. An AI risk firewall is designed to catch all three before execution. (For context on the broader copy-trading category, see what is auto copy trading.)
The architecture in one diagram
publisher account β signal webhook β Foxy AI audit β your account
(OKX/Bybit) (1-3ms latency) (~800ms scoring) (order routes)
β
βΌ
BLOCK if score < threshold
(your envelope, your call)
The audit is a synchronous gate. The signal does not reach your account until Foxy AI has scored it. If the score is below your declared threshold, the signal is logged but not executed.
What Foxy AI reads
The score is a weighted aggregate of about 12 inputs. The five that matter most:
1. Order-book depth. Foxy AI pulls L2 from CoinGlass and the venue's own API. If the proposed entry size is more than ~2% of immediate liquidity, the trade gets a slippage penalty. This is what blocks "size 3Γ normal" entries that signal desperation.
2. Funding rates. Perp funding tells you what the rest of the market is doing. A long position taken at -150% funding is paying the shorts every 8 hours. If the publisher is on the wrong side of crowded flow, the score reflects it.
3. Publisher pattern match. Each publisher has a baseline in Foxy AI's model β typical position size, typical leverage, typical hold time, typical entry vs. recent price. Signals that deviate >2Ο from the publisher's own baseline get flagged. This is what catches revenge trades and leverage drift.
4. News and sentiment. A real-time scan of the last 30 minutes of news for the asset and its sector. Signals that fight known news flow are penalized. (Long ETH 30 minutes after a known L2 exploit gets a penalty.)
5. Whale flow. Large positions opened on Hyperliquid and the major venues are tracked separately. If retail signal X is moving the opposite direction from concentrated whale flow, that's a yellow flag β not blocking on its own, but contributing to the aggregate.
The remaining inputs (volatility regime, FOMC/CPI proximity, asset correlation in the user's existing portfolio, time-of-day liquidity, etc.) refine the score but rarely flip the decision.
What the score actually looks like
Foxy AI returns a 0β100 score per signal, plus a small JSON payload with the contributing factors:
{
"score": 72,
"decision": "ALLOW",
"factors": [
{ "name": "publisher_pattern", "weight": 0.30, "value": 0.85 },
{ "name": "orderbook_depth", "weight": 0.20, "value": 0.78 },
{ "name": "funding_rate", "weight": 0.15, "value": 0.65 },
{ "name": "news_sentiment", "weight": 0.15, "value": 0.70 },
{ "name": "whale_flow", "weight": 0.10, "value": 0.60 },
{ "name": "other", "weight": 0.10, "value": 0.55 }
]
}
The decision isn't binary. It's a number compared against the threshold you set. Conservative users run at 65; aggressive users run at 40. The default for new accounts is 50.
A trade scoring 28/100 is not "the AI thinks this trade will lose money." It's "the AI sees enough red flags that this doesn't pass our risk envelope at the threshold the user set." Many of those blocked trades will print profit. The firewall is about risk-adjusted return, not absolute return β see also AI portfolio management for crypto.
Worked example: the 03:00 GMT revenge trade
A real pattern Foxy AI was designed to catch.
A publisher opens a 5Γ long ETH at 02:40 GMT. Stop hits at 02:55. Twelve minutes later β 03:07 GMT β they open a new position: 25Γ long ETH, position size 4Γ their normal.
What 2010-era copy trading does. Mirrors the trade. Your account is now 25Γ long ETH at 4Γ normal size, taken in illiquid Asian-session hours, by a publisher who just took a loss.
What Foxy AI does. Three flags fire:
- Publisher pattern: position size 4Γ baseline, leverage 5Γ baseline β 0.20 (very abnormal)
- Time-of-day liquidity: 03:00 GMT is the thinnest hour β 0.45
- Recency to prior loss: 12 minutes β 0.30 (revenge-trade signal)
Aggregate score: 31/100. At the default threshold of 50, the trade is blocked. Logged for the publisher's record. Your account stays flat. The publisher's next trade β taken hours later, at normal size β passes the audit normally.
This is the boring, quiet, valuable thing the firewall does.
What it does not do
Three honest limitations:
It doesn't predict winners. Foxy AI is calibrated against risk patterns, not return patterns. Many of its blocked trades would have been profitable. The product is designed to reduce catastrophic loss, not maximize upside.
It doesn't catch coordinated publisher fraud. If a publisher has been running a normal-looking pattern for three months specifically to build a low risk score and then makes one big anomalous bet, the system flags the anomaly β but only on the anomalous trade. It doesn't detect long-game manipulation.
It doesn't replace your risk envelope. Foxy AI is one gate. You still need to set max-drawdown, max-leverage, and max-position-size caps on your own account. The firewall and the envelope work together: the firewall scores individual signals; the envelope caps aggregate exposure.
How to think about it
The mental model that works for users: Foxy AI is a bouncer, not a fortune teller. It looks at the trade walking in the door and decides whether it's the kind of thing that should be walking in. Sometimes the bouncer is wrong and turns away someone fine. Sometimes a bad actor slips through. On aggregate, the bar with the bouncer is a better place to be than the bar without.
If you're considering copy trading and the platform doesn't have a firewall β even a less sophisticated one β you are the last line of defense. That works fine when you're at your phone. It works less well at 03:00 GMT.
BottomUP is a Delaware-incorporated copy-trading marketplace. Foxy AI is BottomUP's risk firewall. Copy-trading functionality is not currently offered to U.S. persons. Past performance is not indicative of future results. Crypto trading carries a high risk of total loss.