DeFi protocols lost $3.8 billion to exploits and attacks in 2023. That was actually an improvement over the $5+ billion lost in 2022, but the number still dwarfs what any traditional financial institution would tolerate as an annual loss figure. What makes the problem stubborn is that most of these losses didn't come from the kind of bugs that code audits are designed to catch. They came from economic design flaws: situations where a protocol's incentive structure, market assumptions, or parameter settings created an exploitable condition under specific market stress.
Smart contract audits are necessary but insufficient. An audit confirms that the code does what the specification says. It doesn't tell you what happens when ETH drops 40% in six hours, gas prices spike to 500 gwei, and three large liquidation bots go offline simultaneously. That kind of failure mode lives in the interaction between the protocol's economic design and real market dynamics, and catching it requires a different class of analysis.
This is where AI-driven risk management for DeFi has developed into a real discipline, not a marketing label.
The risk taxonomy
Before getting into the technology, it's worth being precise about what kinds of risk DeFi protocols face, because different risk types require different mitigation strategies.
Code vulnerabilities are the most understood category. Reentrancy attacks, integer overflows, flash loan exploits that take advantage of unprotected function calls. These are software bugs, and the traditional defense is a code audit performed by firms like Trail of Bits, OpenZeppelin, or Consensys Diligence. Static analysis tools and formal verification methods have improved substantially over the past three years. Code-level exploits still happen, but they're becoming less common relative to other attack vectors.
Economic design flaws are the category that AI risk management addresses most directly. Oracle manipulation attacks, where an attacker distorts the price feed that a lending protocol uses to determine collateral values, and then borrows against inflated collateral. Governance attacks, where someone acquires enough voting power to push through a malicious parameter change. Liquidation cascades, where a market downturn triggers a wave of liquidations that themselves push prices lower, triggering more liquidations in a feedback loop. These aren't code bugs; they're emergent behaviors of complex systems under stress.
Operational risks sit in a third category: admin key compromises, bridge exploits, and social engineering attacks against multisig holders. These are infrastructure and process failures rather than protocol design failures, and they require different mitigations (hardware wallets, timelocks, distributed key management) that fall outside the scope of economic simulation.
The first category gets audited. The third category gets addressed through operational security practices. The second category is where AI simulation and modeling earn their place.
Agent-based simulation as a core technique
The dominant approach to AI-driven DeFi risk management is agent-based simulation, and it's worth understanding the mechanics because the technique is more rigorous than it might sound.
The process starts with building a simulated environment that mirrors the protocol's smart contract logic. Not a simplified model of the protocol; an actual representation of its state transitions, parameter settings, and interaction patterns. Into this environment, the platform deploys thousands of autonomous agents, each programmed to represent a different user archetype. Some agents are arbitrageurs who will exploit any price discrepancy between venues. Some are liquidators who monitor undercollateralized positions and compete to liquidate them. Some are whale depositors who move large amounts of capital in and out based on yield differentials. Some are adversarial agents specifically designed to find profitable attack strategies.
The simulation then runs millions of scenarios, each with different market conditions. What happens if ETH drops 50% in an hour? What if the USDC peg wobbles to $0.97? What if the Chainlink oracle delays by 30 seconds during a period of high volatility? What if gas prices make liquidation transactions uneconomical? For each scenario, the system measures protocol health metrics: the amount of bad debt generated, liquidation efficiency (what percentage of undercollateralized positions were actually liquidated), utilization rates, and whether any agent found a profitable exploit.
The output isn't a single risk score. It's a distribution of outcomes across market conditions, which can be used to set protocol parameters (collateral factors, liquidation bonuses, borrow caps, interest rate curve shapes) at levels that keep the protocol solvent across the tested range of conditions.
I want to be clear about what this is and isn't. It's a form of stress testing adapted for on-chain protocols. It isn't a guarantee against novel attack vectors, and it's only as good as the scenarios and agent behaviors the modelers design. But it represents a substantial upgrade over the alternative, which is setting parameters based on intuition and adjusting them reactively after losses occur.
The companies doing the work
Four firms have established meaningful positions in DeFi risk management, though they approach the problem from different angles.
Gauntlet is, by most measures, the leading dedicated DeFi risk platform. The company runs agent-based simulations against protocol smart contract logic and has become the risk manager of record for several of the largest DeFi protocols, including Aave, Compound, and MakerDAO (now rebranded as Sky). Gauntlet's workflow is distinctive: they analyze protocol parameters, run simulations to determine safe ranges, and then submit their recommendations as on-chain governance proposals that token holders vote on. This creates an unusual dynamic where a risk management firm's analysis is publicly visible and subject to community approval rather than being implemented behind closed doors. Gauntlet has recently begun expanding beyond DeFi into traditional finance risk modeling, which suggests the simulation techniques they've developed on-chain have applicability to conventional portfolio risk.
Chaos Labs competes directly with Gauntlet in the economic security space. They've built a simulation engine focused on modeling protocol behavior under extreme market conditions and work with Aave, dYdX, and Osmosis among others. Chaos Labs also provides real-time monitoring and alerting, which addresses a gap in the pure simulation approach: simulations tell you how to set parameters before stress occurs, but real-time monitoring tells you when actual market conditions are approaching the boundaries of your tested scenarios. The combination of pre-deployment simulation and live monitoring creates a more complete risk management practice.
Risk Harbor takes a fundamentally different approach. Rather than analyzing or simulating risk, Risk Harbor provides parametric coverage, which is closer to insurance than risk management. Their products pay out automatically when specific on-chain events occur: a smart contract exploit that drains funds, a stablecoin depeg below a defined threshold. There's no claims process and no discretionary assessment; the trigger conditions are defined in advance, and payouts execute programmatically. This is risk transfer rather than risk analysis, and it serves a different function in the stack. A protocol might use Gauntlet for parameter optimization and Risk Harbor for residual risk coverage.
Chainalysis operates at a broader scope. The firm is primarily known for blockchain analytics and transaction monitoring, but their risk scoring of DeFi protocols and wallet addresses feeds into the risk management ecosystem. When a traditional financial institution evaluates exposure to DeFi (whether through custody, lending, or direct protocol interaction), Chainalysis risk scores help quantify the compliance and counterparty risk associated with specific protocols and addresses.
For a full view of companies operating at the intersection of crypto and risk management, the AIFI crypto segment directory tracks firms across DeFi infrastructure, analytics, and security.
Where on-chain AI agents fit
There's an emerging intersection between DeFi risk management and autonomous AI agents that I think deserves attention, even though it's still early.
As DeFi protocols grow in complexity and total value locked (TVL), more operational functions are being handled by autonomous agents rather than human operators. Liquidation bots have existed for years, but the newer generation of on-chain agents handles more complex tasks: rebalancing yield strategies across multiple protocols, executing governance votes based on predefined criteria, managing treasury diversification, and monitoring cross-chain bridge flows.
These agents need real-time risk assessment to operate safely. An autonomous agent managing a $50 million yield strategy can't wait for a weekly risk report; it needs to evaluate protocol health, market conditions, and exposure concentrations on a per-transaction basis. This is where the simulation platforms and the agentic AI movement converge. The risk models produced by firms like Gauntlet and Chaos Labs could feed directly into agent decision-making frameworks, allowing autonomous capital allocation that respects quantified risk boundaries.
The AIFI agent registry tracks several agents operating in DeFi contexts, and I expect this category to grow as the infrastructure for on-chain agent coordination matures. The EIP-8004 standard for agent registration is one signal that the ecosystem is moving toward structured, verifiable agent identity, which is a prerequisite for agents to interact with risk management systems in a trustworthy way.
Institutional capital brings institutional expectations
DeFi TVL recovered to over $90 billion in early 2025, and the composition of that capital has shifted. Early DeFi participants were retail users and crypto-native funds with high risk tolerance. The newer cohort includes traditional asset managers, corporate treasuries, and in some jurisdictions, regulated financial institutions testing on-chain operations.
These participants bring expectations shaped by decades of traditional finance risk management. They expect Value-at-Risk (VaR) calculations, stress testing under defined scenarios, documented risk limits, and independent risk assessment. They're accustomed to regulatory frameworks like Basel III that prescribe capital requirements based on quantified risk exposures.
Adapting these frameworks to on-chain protocols isn't straightforward. VaR assumes continuous price distributions, but DeFi markets experience discontinuous jumps (oracle updates, MEV-driven price movements) that violate that assumption. Stress testing requires defining plausible worst-case scenarios, but in DeFi, the interaction between protocols through composability creates emergent risks that don't exist in siloed traditional finance. A position that looks safe in isolation might become dangerous when a lending protocol's liquidation cascade affects the DEX where the collateral asset trades.
The simulation-based approach that Gauntlet and Chaos Labs use is, in my view, better suited to these dynamics than straight adaptations of traditional risk models. Agent-based simulation can capture the composability effects and nonlinear feedback loops that characterize DeFi. But the tooling needs to mature, and the outputs need to be presented in formats that traditional risk managers can interpret and act on.
Regulatory context remains fragmented
DeFi risk management doesn't operate in a regulatory vacuum, though the rules are still forming. The SEC's position on DeFi has focused primarily on whether protocols constitute securities offerings. The CFTC has taken enforcement actions against DeFi protocols offering derivatives. Neither agency has published specific guidance on risk management standards for DeFi protocols.
In the EU, the Markets in Crypto-Assets (MiCA) regulation establishes requirements for crypto-asset service providers, but its applicability to decentralized protocols (as opposed to centralized exchanges) is ambiguous. The UK's FCA has taken a cautious approach, focusing on consumer protection rather than protocol-level risk standards.
What seems likely, though I'd hedge this view, is that regulatory expectations for DeFi risk management will increase as institutional participation grows. Protocols that can demonstrate quantified, independently validated risk management practices will have an advantage in attracting institutional capital. Protocols that can't will find themselves limited to retail participation, which may be fine for some but constrains growth for others.
Looking ahead with appropriate caution
The DeFi risk management space has matured from "we did a code audit" to "we run continuous agent-based simulations and submit parameter recommendations through governance." That's genuine progress. The losses in 2023, while still large in absolute terms, came down from 2022, and the protocols that employed dedicated risk management firms generally performed better during market stress events than those that didn't.
I don't want to overstate the case. The simulation models are imperfect. Novel attack vectors can fall outside the tested scenario space. The economic incentives that drive adversarial behavior in DeFi are strong enough that attackers will continue finding blind spots. And the regulatory environment could shift in ways that reshape or constrain DeFi activity regardless of how good the risk management becomes.
What I can say with more confidence is that the demand signal is clear. Protocols competing for capital, particularly institutional capital, need to demonstrate rigorous risk management. The firms building these capabilities have defensible positions based on proprietary simulation infrastructure, protocol relationships, and accumulated modeling expertise. As DeFi TVL grows and the stakes increase, the gap between protocols with and without production-grade risk management will widen. That makes this one of the more durable sub-sectors within the broader crypto and AI intersection.