SEC subpoenas banks over AI hedge fund Situational Awareness collapse

The U.S. Securities and Exchange Commission has subpoenaed multiple Wall Street banks as part of an investigation into a firm called Situational Awareness, an artificial intelligence-focused hedge fund that nearly collapsed in July before securing a rescue deal with Citadel. The SEC’s inquiry underscores growing regulatory scrutiny of how financial institutions interact with AI-driven investment strategies, particularly when those strategies involve opaque, high-risk models that can destabilize markets or expose counterparties to sudden losses.

Situational Awareness, founded by Leopold Aschenbrenner, a former OpenAI researcher, operates at the intersection of AI and quantitative finance. The fund’s near-collapse in July was triggered by a sudden, sharp drawdown, which sources attribute to the failure of its proprietary AI-driven trading models to anticipate market movements. The firm’s reliance on large language models and reinforcement learning to generate trades, rather than traditional fundamental analysis, created a black-box risk for its lenders and prime brokers. When the models produced a series of wrong-way bets, margin calls piled up, and Situational Awareness came close to insolvency before Citadel stepped in with a capital injection and restructuring agreement.

The SEC’s subpoenas are likely focused on what the banks knew about Situational Awareness’s operations and risk controls before the July crisis. Regulators are probing whether the banks, as prime brokers and lenders, adequately vetted the fund’s AI models or were misled about the sophistication of its risk management. The investigation also raises questions about how banks should classify and monitor exposure to AI-driven funds that may not fit traditional risk categories, such as those based on volatility or correlation.

This case highlights a broader tension in financial markets: the push by quantitative funds to harness AI’s predictive power versus the opaque nature of those models. Unlike systematic trend-following or mean-reversion strategies, which have decades of track records and known risk profiles, AI-based funds often rely on neural networks whose decision-making processes are not fully interpretable even to their creators. That creates a fundamental problem for risk managers, who cannot easily stress-test or audit the models, and for regulators, who must ensure that systemically important banks are not unknowingly underwriting unquantifiable risks.

The near-collapse of Situational Awareness is also a cautionary tale for the AI industry’s foray into capital markets. Aschenbrenner, a well-known figure in AI safety research, had argued that AI models could uncover patterns in market data that human analysts and conventional quant models miss. But the July episode shows that the same models can also fail spectacularly when market regimes shift, precisely because they are trained on historical data that may not reflect future conditions. The Citadel rescue, while stabilizing Situational Awareness in the short term, introduces its own conflicts: Citadel now has a window into the fund’s models and trading strategies, potentially giving it an edge in the same markets where it competes.

For investors, the episode is a reminder that AI is not a magic wand. The hype around generative AI and large language models has spilled into finance, but the fundamental challenges of risk, leverage, and liquidity remain unchanged. The SEC’s subpoenas are a signal that regulators intend to scrutinize this intersection closely, and that banks may need to demand more transparency from AI fund managers before extending credit. Meanwhile, the industry will be watching to see whether Situational Awareness can rebuild its track record under Citadel’s oversight, or whether it becomes the first prominent casualty of the AI bubble in markets.