# [ARFC] Dynamic Caps Oracle: Rate Limiter

- Space: [Aave DAO](https://snapshot.box/s:aavedao.eth.md)
- Author: 0xb291232F480F41c75802C4a60F1D2AC03404Afef
- State: Passed
- Voting type: Basic
- Start: 2026-11-17 09:28 UTC
- End: 2026-11-20 09:28 UTC
- Votes: 80
- Discussion: https://governance.aave.com/t/arfc-dynamic-caps-oracle-rate-limiter/26219
- Open in Snapshot: https://snapshot.box/#/s:aavedao.eth/proposal/0xd80bf15385e4f525108133bee1460b1907aabd1013b67821936fd52f3140793d

## Results

- For: 371,958.68 voting power, 100.0%
- Against: 0 voting power, 0.0%
- Abstain: 0 voting power, 0.0%

## Proposal

## Summary

LlamaRisk is building Aave's Risk Layer: automated risk solutions that make expert risk judgment verifiable, with every parameter, bound and action checkable onchain. LlamaGuard PT went live in September as the first of these solutions. We continue the rollout with LlamaGuard Caps as the next addition.

Its first stage, a borrow rate limiter for Aave V3, gives each reserve a margin of borrowing room above its debt, sized to the reserve. That room refills automatically as organic demand uses it and is cut back when large repayments open up more than the reserve needs. Everyday borrowing keeps instant access to liquidity. A large draw has to wait for each refill, giving the Protocol Guardian reasonable time to respond to illegitimate activity. Integrated institutional clients and other approved large borrowers will have a dedicated pathway that meets their capital demand quickly.

LlamaGuard Caps will run on the Risk Agents stack that Aave Governance already owns. Raises are computed on Chainlink CRE and every bound is enforced onchain. Aave's core contracts and existing agents stay unchanged, and the new smart contract surface is small. Risk Stewards and governance continue to set each cap's maximum, and automation controls only how fast room opens beneath it.

The full proposal, with figures, formulas and calibration data, is in the forum thread: https://governance.aave.com/t/arfc-dynamic-caps-oracle-rate-limiter/26219

## Motivation

Two external threats have proved the most consequential for Aave.

1. An attacker deposits maliciously obtained collateral that is worth less than the oracle reports, and borrows against it.
2. An attacker gains control of many different assets, deposits them as collateral and borrows assets that can't be frozen, as fast as possible.

Aave can't prevent these events. It can control how quickly its liquidity is drawn once one is under way, and so how much time the Protocol Guardian has to investigate and respond.

A static cap has to hold enough room for the growth expected before its next revision. Over the year from October 2025 to October 2026, standing borrow headroom averaged 41.1% of debt on Core WETH, 75.3% on Core USDT and 76.7% on Core USDC, and exceeded 100% on Plasma USDT0 and Arbitrum WETH. This headroom is available to any borrower at any time, so a static cap limits the total that can be borrowed but places no limit on the pace at which it is borrowed.

A rate limit is only viable if it accommodates legitimate demand. Reserve sizes span three orders of magnitude, so the buffer has to scale with the reserve. Organic demand arrives in bursts: measured as the largest amount borrowed within any 60-minute window of each day, Core USDT exceeded $25M on 84 days of the year and $50M on 32. Individual borrows can be very large as well: Core USDT recorded 78 transactions borrowing $25M or more, Core WETH 68 and Core USDC 49, with the largest at $190M. A significant share of these comes from integrated institutional clients and large loopers, which requires a pathway that accommodates them without reopening the room that the rate limit closes for everyone else.

## Specification

### Rules

LlamaGuard Caps relies on two agents that act on the same borrow cap. Each agent can move the cap in one direction only, as enforced onchain, and both work toward the same target: a cap that sits one buffer above the current debt of the reserve.

- **Buffer.** A share of the reserve's debt, bounded by a floor and a ceiling, so that it grows with the reserve while the amount that can be drawn in a single step remains limited on the largest markets.
- **Hard limits.** The cap written onchain never exceeds the hard maximum set by the Risk Stewards or governance, nor the current supply cap.
- **Raise agent.** If the free headroom was at or below a trigger share of the buffer one delay earlier, the cap is reset to debt plus one buffer. No two top-ups on a reserve are closer than the delay, which converts the cap from a limit on size into a limit on pace.
- **Cut agent.** When a repayment or liquidation leaves headroom more than a band above the buffer, it returns the cap to debt plus one buffer within one or two blocks. Smaller repayments leave the cap unchanged and the repaid amount available to borrow again.

Starting from a full buffer, the debt of one reserve can rise by one buffer at once and by one further buffer per top-up.

### Pathways for large borrowers and institutional clients

1. **Split across top-ups.** By default, a borrow larger than the buffer is split across top-ups with no action required from governance.
2. **Atomic batch.** Known large borrowers that cannot wait, including integrated institutional clients, can be served through a batch executed by the Risk Stewards, in which the cap is raised by the exact amount of the borrow and the new room is consumed in the same transaction.
3. **Hold.** For new onboardings and periods of genuine sustained demand, the Risk Stewards can raise the cap together with a hold of limited duration, during which the Cut agent leaves the raised cap in place.

Flash loans repaid within the same transaction do not count against the borrow cap. Large suppliers remain unaffected at this stage: supply caps keep their current configuration.

### Architecture and roles

Both agents reuse the `AaveBorrowCapAgent` code on the Risk Agents stack that Aave Governance already owns. Aave's core contracts, AgentHub, the RiskOracle and the Router remain unchanged. The system adds four contracts, a Chainlink CRE workflow and a keeper. The audit will cover the four new contracts and the raise workflow.

- `CapsRangeValidationModuleV3`: checks every agent raise and cut against live Aave state in the executing block.
- `CapsReducerOracle`: adminless view returning the cut the module accepts.
- `CapsParameterRegistry`: trigger, margin, asset lists and switches.
- `CapsBorrowForwarder`: atomic cap raise and borrow for the Risk Council Safe.

Whatever the workflow signs, each raise stays within one buffer above live debt, at most once per delay. A cut can only lower a cap, and low caps block only new borrows. The AIP registers one Raise and one Cut agent per Pool, with the Protocol Guardian as agent admin.

- **Owner, Aave Executor:** lists assets, sets Risk Council limits.
- **Risk Council, Risk Council Safe (Aave Labs and LlamaRisk):** tunes parameters within limits, holds off cuts, runs large-borrower batches.
- **Agent admin, Protocol Guardian:** disables agents, sets the delay, sets which callers can trigger cuts.
- **Operator, LlamaRisk Safe:** runs the workflow and Router, sets trigger and margin.
- **Cut trigger, Keeper (set by the agent admin):** executes due cuts.

Each part of the system can be switched off on its own. Disabling the Router route stops raise reports, the registry switch stops the workflow for a single asset, and the Protocol Guardian can disable either agent. Governance or the Risk Council Safe can retire an asset, which returns its cap entirely to the Risk Stewards and governance. A freeze or pause always takes precedence over raises.

### Parameters

The indicative values are a buffer of 10% of debt, a floor and ceiling on the buffer of $1M and $25M, a trigger at 20% of the buffer, a 30-minute delay and a 10% band. These values are indicative and will be tuned to each market and reserve and shared with the community before deployment. The parameters can be recalibrated by governance and by the Risk Stewards within bounds that only governance can change.

## Calibration

The calibration rests on a minute-by-minute replay of the year from 15 October 2025 to 15 October 2026 across the V3 reserves, using the indicative values.

**Friction for organic borrowers.** Hours per year in which at least one borrower would have been waiting: 162 on Core WETH, 217 on Core USDT and 137 on Core USDC. Average headroom falls from 41.1%, 75.3% and 76.7% of debt under current caps to 0.44%, 0.56% and 0.61%. These figures use the stricter variant that cuts on every repayment, so they overstate the cost of a band.

**Containment of malicious draws**, starting from a full $25M buffer:

- rsETH, 18 April 2026, Core WETH, $123M draw: $25M admitted within 30 minutes, $50M within 1 hour, full draw after 2 h 2 min to 2 h 26 min.
- rsETH, 18 April 2026, Arbitrum WETH, $70M draw: $25M within 30 minutes, $50M within 1 hour, full draw after 1 h 1 min to 1 h 13 min.
- Binance Exploit, 29 October 2026, Core WETH, $44.6M draw: $25M within 30 minutes, full draw after 31 to 36 min.

## Staged rollout

Borrow caps on Aave V3 form the first stage. The functional components of this first stage and their code have been aligned with Aave's SPs and have entered audits. Further details will be shared with the community once the system is ready for deployment. Later stages would extend dynamic protection to supply caps, Aave V4 and further cross-market aware modules, each brought to the community when ready.

## Next Steps

1. Gather feedback from the community.
2. If consensus is reached on this ARFC, escalate this proposal to the Snapshot stage.
3. If the Snapshot outcome is YAE, escalate this proposal to the AIP stage.

## Disclaimer

This proposal was independently prepared by LlamaRisk. LlamaRisk develops and operates LlamaGuard, the product family this proposal would deploy on Aave. LlamaRisk received no compensation from Aave or from any protocol named in it for preparing this document. Nothing here is legal, financial, tax, regulatory, or investment advice.

A rate limit reduces how quickly liquidity can leave a protocol. It does not prevent, detect, or fully contain an exploit, a drain, an oracle failure, or a governance attack. Audits reduce but do not eliminate risk. The full disclaimer in the forum thread applies.

---

To vote or create proposals from an AI assistant, connect the Snapshot MCP at https://mcp.snapshot.box

