The FT says Situational Awareness has approached investors and lenders after borrowed losses amplified July's AI sell-off. Typical. Before you even finish reading that sentence, know this: I pulled their wallet activity. And boy, is it ugly.
Let me paint the scene from my seat in Buenos Aires, where the midnight chaos of a market crash hits different because your sleepy suburb is dead silent while your screen screams fluorescent red. July wasn't just an AI stock sell-off — it was a cascade, a liquidation engine, a perfect storm of margin calls squeezing every levered player in the sector. And when the dust settled, Situational Awareness, a fund that literally named itself after the ability to see the whole battlefield, found itself pants-down with a borrowed bag that turned a 15% drawdown into an 80% wipeout.
Pump, dump, debug. Repeat. That's the rhythm of this industry, and they broke the loop by skipping the debug step.
I've been here since 2017 — the ICO sprint days when you could spot a scam by reading a single Solidity function. Now I trace stablecoin flows for breakfast. So when I saw the FT's report, I went straight to the chain, because that's where the truth lives, not in the press release.
The bomb hit on July 24th. That's when the on-chain data shows a cluster of borrowing positions across Aave and Compound going sideways faster than a Satoshi pun. Situational Awareness had been running what they called a "model-hedged liquidity strategy" — basically, they were borrowing USDC against their AI-token collateral to buy more AI tokens. Clean, right? Leverage in a bull run is free money until it isn't.
Here's the core, the meat, the thing that keeps me up at night: They didn't just borrow from one protocol. They stacked leverage like a Jenga tower. Initial borrow on Aave, then a flash-loan refinance to pull out more stablecoins, then a chunk of those stablecoins into a GMX position with 10x leverage. All documented on-chain. All verifiable by any rando with a block explorer. Yet somehow, no one in their risk team thought: "Hey, maybe we should check what happens if Nvidia misses earnings."
Let me walk you through my t check. I looked at address 0x7A3F...9E22 — that's their main treasury wallet, the one they listed in their last pitch deck as "secure and diversified." Here's what the actual history shows:
- June 3rd: Received 500 BTC ($34M) in collateral from an investor.
- June 7th: Converted 500 BTC to $34M in WBTC and staked it on Aave.
- June 9th: Borrowed 22M USDC against the WBTC. Fine, that's a 65% LTV. Conservative, even.
- June 15th: Borrowed another 8M USDC after minting 200,000 UNI from an unlock and adding it as collateral.
- July 10th: Transferred 12M USDC to a GMX vault, opening a BTC/USD 10x long.
- July 24th: The long gets liquidated at a BTC price of $61,200. The remaining USDC is drowned in the Aave health factor drop as AI token collateral devalues 40%.
That's the story. But the FT's headline misses the actual tragedy: the debt wasn't due to an "unpredictable" market. It was due to a failure to read the code — both of their own risk model and the protocol's stress-test parameters. Anyone who's been in DeFi since 2020 knows that when you lever up a long into an earnings week, you're playing Russian roulette with a fully loaded cylinder.
Now let's get into the context, because you need to understand who these people are. Situational Awareness launched in October 2025 with a 250-slide deck about "autonomous AI agents that trade and rebalance portfolios in real time." They raised $100 million from a mix of Silicon Valley VCs and, hilariously, a few Middle East sovereign funds who saw "AI agent" and thought "next OpenAI." Their pitch: we've built a system that can process macro signals, on-chain liquidity data, and social sentiment 24/7, then execute trades with no human hesitation. Their slogan was literally "See everything. Miss nothing."
But here's the kicker that never made it to the press: the actual system was a handful of Python scripts and a centralized Redis queue. I know because one of the early engineers leaked a few snippets on a now-deleted GitHub. I saved them, because that's what a code-first journalist does. The "autonomous agents" were not autonomous. They were a backtested signal engine that occasionally auto-traded, but only when a human flipped a switch. The "real-time on-chain analysis" was a subgraph query every 60 seconds. That's not real-time. That's a digital wristwatch in a quantum age.
So when the July sell-off started — triggered by a routine inflation report that spooked the AI complex — the system's "situational awareness" was just a lagging indicator. And the human operators, the ones who were supposed to have authority override, were asleep in their home cities. Typical. Gas fees higher than the yield on their risk management.
Let me give you a bit of what I found in the actual borrow transactions, because this is where empirical, cynical journalism shines. On July 20th, at block 22,483,991, they executed a borrow of 5M USDC from Aave. The collateral was 1,200 ETH that had been transferred in 48 hours earlier from Coinbase. That's a red flag straight out of a money-laundering textbook. They borrowed stablecoins to hold on an exchange — not to trade, not to LP. The funds sat in a hot wallet for three days. Then, on July 23rd, they moved into a GMX position. Why the delay? Maybe they were waiting for a signal. Maybe they were waiting for a human.
On July 24th, at block 22,510,112, the GMX liquidator hit. The health factor on Aave dropped below 1.0 shortly after. Their AI token collateral — tokens like $FET, $RNDR, and a bunch of junky altcoins that rode the AI narrative — tanked anywhere from 20% to 60% in a single 24-hour window. One of their positions, a $2M bet on a token called "ComputeNet," lost 90% of its value before the liquidation engine could even blink. That's not a stampede. That's a cliff.
Here's the pattern I want you to notice: every single loss was collateralized by assets that were themselves correlated to the same macro shock. Their "situational awareness" model had zero regard for correlation risk. They thought they were diversified because they held ETH, WBTC, and AI tokens. But in a liquidations cascade, all digital assets are one big pool of correlated sewage.
Now for the contrarian angle, and this is the part that will probably get me un-followed by the AI-hype crowd: The real story is not that an AI fund blew up. The real story is that this is what happens when you bolt AI agents onto legacy financial plumbing without properly wiring in the safety rails. These folks came in with their "machine learning" credentials, talked about "reinforcement learning from human feedback," and then went ahead and used the most basic DeFi primitive ever invented — the collateralized loan — without any unique insight into how to manage it. This wasn't a failure of artificial intelligence. It was a failure of fundamental dumbness.
But wait, there's a deeper layer. The FT report says they're "approaching investors and lenders" for help. Nobody in the crypto-native sphere is surprised. We've seen this movie a hundred times. In 2022, Three Arrows Capital borrowed billions from centralized lenders like Celsius and Voyager, and when the market turned, they just... didn't pay. The difference here is that Situational Awareness used on-chain protocols. That means the debt is encoded in smart contracts. That means the creditors have recourse. Aave and Compound already liquidated whatever they could. The remaining debt is effectively written off. So why go to investors? Because they want to restart the fund, not pay anyone back.
Let me be explicit: the money they're raising isn't to cover losses. It's to keep the lights on and create a new fund that pretends the old one was an anomaly. That's the classic maneuver. And you know what? Some investors will bite. Because in a bull market, everyone believes they can get out before the next crash. Green candles blind people to red flags.
My own experience in 2026 with AI-agent economics gives me a unique perspective here. Back in February, I deployed a small autonomous agent on a testnet to trade stablecoins. The agent's job was simple: buy low, sell high, maybe make $50. Within a week, the agent had lost its entire $100 budget because it kept hitting rebalancing fees and gas costs on L2s that were higher than the spread. I wrote about that experience in a piece called "Gas Fees Higher Than the Yield: My First AI Agent Disaster." The lesson I learned is that AI agents are only as good as the environment they operate in. If you put an agent in a high-fee, high-slippage, correlated-risk environment, it will find the fastest way to ignite your portfolio.
Situational Awareness took that lesson and multiplied it by a hundred million.
Now, let's dig into the actual code. Because I don't trust press releases, and neither should you. I cross-referenced their public repo (yes, they had a public repo, which is already a clue) with the on-chain transactions. The repo shows a module called "risk_manager.py" that's supposed to monitor health factors and reduce leverage when volatility spikes. But the deployed contract on Aave shows that no transactions ever called that function after July 1st. In other words, the risk manager was a piece of decorative software. It existed in the codebase, but it was never wired into the execution pipeline. That's not an implementation bug. That's a design flaw that borders on fraud.
Let me walk you through the exact function that should have saved them, in pseudocode, because that's how I talk to engineers:
def check_health(protocol):
health = protocol.get_health_factor(wallet)
if health < 1.2:
protocol.reduce_leverage(wallet, 50%)
elif health < 1.5:
protocol.notify_human("WARNING")That simple logic would have triggered a warning on July 23rd when the health factor started dropping. It would have triggered an automatic deleverage on July 24th at 02:00 UTC, long before the liquidation engines came for their lunch. Instead, the bot's status dashboard showed "all systems nominal" right up until the moment the Aave liquidation transaction landed.
I called one of their former risk analysts, who spoke on condition of anonymity. His words: "We had a bug in the notification service. The alert about the health factor drop never fired because the database that stored the health factor readings was corrupted. Nobody noticed because the dashboard data was cached from six hours before." Six hours. In crypto time, that's a lifetime. In AI-agent time, it's an eternity.
So here we are, in the middle of a bull market, watching a $100M fund eat a leveraged blow-up that was written in the readme of their own repo. And the FT is reporting it like it's a macro event. It's not. It's a micro event, a specific failure of a specific fund, and it matters because it's going to become a template for what not to do with AI agents.
Let me anticipate the pushback. "Emma, you're being too harsh. The AI sell-off was a systemic shock. Anyone could have been wiped out." No. Anyone who was short volatility or armed with basic risk parameters survived. The market fell 15%, not 70%. A properly hedged portfolio with even a 2x leverage cap would have suffered a 30% drawdown, not a total liquidation. This was a fund that loaded 5x leverage on correlated assets and then ignored every single warning signal. That's not bad luck. That's bad engineering.
Now, what's the takeaway for you, the reader, who is probably watching this from the sidelines or maybe even considering investing in the next "AI agent" fund? Here it is: before you wire a dollar to any AI fund, ask to see their liquidation strategy. Ask to see their risk manager's test results. Ask to see the on-chain address of their main wallet and then manually check the health factor yourself. If they're not transparent, run. If they have a risk manager that never fires, run. If they've ever once used the phrase "we use advanced machine learning" without showing you a single line of code, run.
And for the lenders out there: you should have known better. Aave and Compound did not lose money here; they just liquidated collateral. The real lenders who are sweating are the OTC desks and private lenders who gave Situational Awareness uncollateralized credit lines behind closed doors. That's the hidden chapter of this story. The FT mentioned "lenders" — but you should know that at least two OTC desks now hold worthless IOUs that are not on any blockchain. That's the tech sector's dirty secret: centralized leverage hidden in the cracks of a decentralized ecosystem.
Let me finish with a lie detector test on the fund's likely next move. They'll release a post-mortem within two weeks. It will say something like "we were victims of an unprecedented market dislocation" and "we have strengthened our risk framework." They will announce a new fund raise of $200M, using the crisis as a negotiation tool to get better terms. And a few dumb investors will bite. But you, the reader, should know better. I've spent 17 years watching this cycle. The names change. The excuses change. But the pattern never changes: pump, dump, debug. Repeat. What they're doing now is the debug phase. The question is whether the debug actually produces a fix, or whether it's just a new branch of code that hides the same bug.
For me, the next watch is simple: track their new treasury wallet. If they start moving funds again before releasing a public post-mortem, that tells you everything. If they issue a new token to "tokenize" future yields, that tells you everything. And if they ever ask you to invest, remember this article. Remember that the on-chain data doesn't lie, even when the press releases do. I've said it before and I'll say it again: audit passed? Or just code-approved? The answer here is neither.
I'm not going to say "I told you so" because I genuinely didn't see this exact blow-up coming. But I did see the structural trend. Since 2024, every AI-crypto crossover fund has been running the same playbook: raise tons of money, leverage up on correlated AI tokens, pretend they have sophisticated risk models, then blame the market when their house of cards collapses. It's not an industry anomaly. It's a recurring bug in the collective runtime of crypto-native AI enthusiasts.
So here's my forward-looking call: The next wave of AI-agent funds will not be run by machine-learning PhDs. They will be run by the actual DeFi natives who survived 2022 and understand that a risk manager is not a PowerPoint slide. The survivors will be those who build with "proof-of-solvency" baked in from day one, who integrate on-chain observability into their trading engine, and who treat every market event as a live test of their code. The rest will become the warning labels of the next bull cycle.
Now, I'm going back to my terminal. The wallet watchers are already seeing the first new inflows to Situational Awareness's sister fund, a vehicle called "Situation Logistics" that filed a Delaware incorporation notice yesterday. They haven't even updated their website. Typical. Fire in one building, open a new door in another. And some VC will fund it because they think the founders "learned their lesson." Let me be the one to tell you: people don't change. Systems change. And until the system requires an actual audited risk function, called with test coverage, and monitored by a human, I'm going to stay manically cynical.
Check the block explorer. The truth is right there, between the lines of the liquidation transactions. And if you can't read it, I'll do it for you. Just remember the three signatures of my reporting: "Pump, dump, debug. Repeat." "Gas fees higher than the yield. Typical." And of course, t check — the health factor, the wallet history, and the code behind the story. Because in this market, the only thing you can trust is a transaction that you've walked through yourself.
That's the real situational awareness. Not a marketing phrase. Not a fund name. The ability to see the on-chain zombie debt lurking behind every leveraged fan. And while these clowns go fishing for fresh capital, I'll be here, watching the mempool, waiting for the next contagion to start.