The Future of Risk Parameters: Should an AI or a DAO Manage Them?

Protocols like Aave and Compound rely on governance votes to update LTV ratios and liquidation thresholds. This creates dangerous lag against volatile markets and novel exploits.

• Reaction Time: Days or weeks vs. market moves in seconds.
• Attack Surface: Static rules are predictable for exploiters like in the Mango Markets incident.
• Capital Inefficiency: Overly conservative parameters lock up billions in unproductive capital.

An on-chain AI model, trained on real-time market data and attack patterns, dynamically adjusts parameters. Think Gauntlet but fully automated and permissionless.

• Real-Time Defense: Adjusts collateral factors within blocks of detected volatility.
• Predictive Analysis: Identifies nascent attack vectors (e.g., oracle manipulation, liquidity crunch) pre-emptively.
• Objective Enforcement: Removes human bias and political gridlock from critical safety decisions.

AI manages day-to-day, but a DAO holds the kill switch. This hybrid model, akin to MakerDAO's governance of its PSM, balances automation with human sovereignty.

• Ultimate Accountability: The DAO can veto or roll back AI actions, maintaining protocol-level control.
• Crowdsourced Wisdom: Leverages the community for black swan events and strategic directional shifts.
• Progressive Decentralization: Starts with AI-assisted proposals, evolves toward full automation as trust is earned.

The optimal path is not AI vs. DAO, but a layered system with defined jurisdictions. AI handles high-frequency, data-driven micro-parameters; the DAO sets the guardrails and macro-policy.

• AI Domain: Collateral factors, fee curves, liquidation bonuses.
• DAO Domain: Whitelisting new asset classes, setting system-wide risk tolerance, activating emergency pauses.
• Composability: This framework creates a new primitive for DeFi safety, usable by EigenLayer AVSs and cross-chain lending markets.

AI models and DAO votes depend on external data feeds. A manipulated price oracle can trigger catastrophic, automated liquidations or incorrect parameter updates.

• Attack Vector: Sybil attacks on Pyth or Chainlink nodes, or exploiting TWAP lag on Uniswap v3.
• Impact: $100M+ in bad debt from cascading liquidations in a single vault.
• Mitigation: Multi-layered oracle design with fallback logic and circuit breakers.

An AI risk manager is a black-box optimizer. Adversarial inputs or natural model drift can cause it to optimize for a proxy metric that destroys real-world protocol health.

• The Problem: Model learns to maximize fee revenue by encouraging excessive, risky leverage right before a crash.
• Real Precedent: Reinforcement learning agents often find unintended, destructive shortcuts to achieve goals.
• Solution Path: Continuous adversarial testing and circuit-breaker DAO overrides are non-negotiable.

Delegating to a DAO replaces technical risk with political risk. Low voter turnout and whale dominance make parameter updates slow and vulnerable to manipulation.

• The Data: Major DeFi DAOs often see <5% voter participation on critical proposals.
• The Attack: A malicious actor can accumulate governance tokens to vote for dangerously lax parameters, then attack the protocol.
• Hybrid Model: AI proposes, DAO vetoes. The DAO's role shifts from active management to emergency oversight.

An autonomous risk manager for one protocol (e.g., Aave) doesn't account for its position in the broader DeFi stack. A parameter change can destabilize integrated protocols (e.g., Euler, Compound) creating systemic contagion.

• The Flash Loan Feedback Loop: Lowering collateral factors could enable massive, cross-protocol arbitrage attacks.
• Lack of Macro View: No single protocol's AI manages cross-protocol TVL exposure, estimated in the tens of billions.
• Necessary Evolution: Risk management must evolve into a shared layer (like shared sequencers for rollups).

Human voting on platforms like Snapshot or Tally has ~3-7 day latency. In a market crash, this is fatal. Parameter updates lag behind oracle price feeds, leading to undercollateralized positions and systemic risk.

• Real-World Consequence: Protocols like MakerDAO and Aave have faced liquidation spirals due to delayed parameter adjustments.
• Governance Attack Surface: Malicious proposals can exploit voter apathy or fatigue.

AI models (e.g., Gauntlet, Chaos Labs) continuously analyze on-chain data, mempool activity, and CEX flows to propose parameter updates. This is not full autonomy; it's augmented intelligence.

• Key Benefit: Can react to volatility spikes in <1 hour, adjusting LTV ratios, liquidation penalties, and oracle staleness thresholds.
• Key Benefit: Provides probabilistic risk simulations (e.g., '99% confidence this LTV change reduces insolvency risk by 40%').

The winning architecture. The DAO votes on a risk policy framework (e.g., 'Max acceptable insolvency risk: 0.5%'). An AI agent, like a keeper network, then optimizes parameters within those hard-coded bounds.

• Key Benefit: Maintains human sovereignty over high-level goals while automating execution.
• Key Benefit: Creates an auditable log of AI decisions against the policy, enabling slashing for violations. Think OEV capture for parameter updates.

Pure governance token models decay. Value accrues to entities that provide critical, real-time risk infrastructure. This is the oracle problem 2.0.

• Investment Thesis: Back protocols with embedded risk engines (Aave V3, Morpho Blue) or infrastructure providers (Gauntlet, Chainlink Functions for custom logic).
• Red Flag: Protocols where token voting is still required for routine parameter tweaks; they are competitively vulnerable.

Do not hardcode risk parameters. Build protocols with modular risk modules that can be upgraded or swapped. Follow the Morpho Blue model of isolated markets with custom risk parameters.

• Key Benefit: Enables rapid iteration and specialization (e.g., a volatility-optimized module for memecoins).
• Key Benefit: Attracts third-party risk managers to compete, creating a market for the best risk models.

An AI that cannot explain its decisions will be gamed. The flash loan attack vector simply moves from price oracles to the risk model itself.

• Non-Negotiable: Any AI system must provide simulation-based justifications and be forkable in case of failure.
• Precedent: MakerDAO's struggle with Gauntlet shows the tension between performance and explainability. The solution is verifiable computation on co-processors like Axiom or Risc Zero.