AI already runs the monitoring, the simulations, and a lot of the automation behind on-chain risk. Here is what it still can’t do on its own, and why your yield still rests on human judgment.

Here is the short answer. AI already manages real parts of on-chain risk today.

It runs simulations, watches markets every second of every day, and adjusts protocol settings inside limits that people approve.

What it does not do yet is run the whole risk desk on its own. So the honest take on AI risk management in DeFi is this: yes for the hard, repetitive parts, and not yet as a manager you leave alone.

That gap matters more than the hype suggests, so let me walk through what AI does, where it hits a wall, and what measured on-chain risk really looks like.

The quick version

• AI already runs simulation, monitoring, and bounded automation across DeFi.
• Firms like Gauntlet and Chaos Labs provide much of that risk plumbing.
• Automated systems act only within human-set, governance-approved limits.
• The real limits stay: stale data, models that get gamed, and crises no model has seen.
• Over-automation can spread risk, not contain it. AI augments risk management; it does not replace it.

What “managing on-chain risk” actually means

Managing on-chain risk means keeping a protocol solvent and liquid while markets move.

In plain terms, that is four jobs. Set the rules: collateral ratios, borrow caps, interest curves. Watch everything: positions, liquidity, oracle prices, concentration.

Stress-test: model crashes and attacks before they happen. Act fast: change settings the moment conditions change.

It is constant, data-heavy work, and the cost of getting it wrong is bad debt. No wonder teams want to automate it. It is also why a careless version is dangerous.

Where AI already runs the risk desk

AI already handles the simulation, monitoring, and bounded automation behind DeFi risk, and it is now core infrastructure. A few concrete examples, as of late 2026.

Gauntlet builds simulation-driven risk models and has acted as the economic risk brain for major lending markets, tuning parameters like collateral ratios and caps. It now runs some of the largest curated vaults too.

Chaos Labs runs AI-powered economic and security stress tests. Its risk oracles watch indicators in real time and push parameter changes within governance-approved bounds.

Aave’s renewal with the firm described more than 1,100 parameter updates.

AI-curated vaults from firms like Chaos Labs, Gauntlet, and Sentora now sit behind consumer products, rebalancing capital across Aave, Morpho, and Sky inside set risk targets.

The division of labor is simple. The machine does the speed: monitoring, simulation, anomaly detection, bounded adjustments.

The human keeps the judgment: which scenarios to run, which thresholds matter, which bounds to allow. This is real automated risk management in crypto, not a demo.

The part AI still can’t do

AI cannot be trusted to manage on-chain risk unsupervised, and 2026 showed why.

Start with the limits that never go away. It works inside human bounds, so that judgment stays human.

Models depend on data and oracles, and a stale price turns a smart system into a confident fool.

Models can be gamed by adversaries who know which inputs break them. And tail events escape training, because the crisis that matters most is usually the one the model has never seen.

The year drove this home. A string of lending blowups traced back to hardcoded or stale oracles.

In one case a collateral token was marked on-chain near $1.13 while it traded around $0.63 on the open market.

That is not a model being clever. That is automation acting fast on a wrong number.

The uncomfortable twist: automation can add risk. Fast, automated allocation paired with slow or sloppy controls, or a risk curator that itself fails, can spread a problem instead of containing it. Speed cuts both ways.

Governance-safe automation: humans still set the bounds

The norm today is not autonomous AI. It is governance-safe automation: people set the strategy and the limits, and AI acts inside them.

People decide the risk appetite and approve the guardrails. AI monitors, simulates, and makes bounded changes, often with a validation step before anything goes on-chain. And a human brake always exists.

You keep automation’s speed and keep accountability. Handing a model the power to act, with real money and real adversaries on the other side, should not be done blind.

What measured on-chain risk looks like: a Sky case study

The real answer to on-chain risk is not add AI. It is make the risk measurable. Sky is one of the clearer examples of that in stablecoins.

A quick primer, since the names trip people up. USDS is Sky’s fully backed stablecoin and the way into the system. It is the unit of account, not the yield.

sUSDS is what you hold to earn: you supply USDS, receive sUSDS, and it accrues the Sky Savings Rate over time. It is currently the largest rate-bearing stablecoin by supply.

The Sky Savings Rate is the governance-set yield behind sUSDS, funded by revenue from the Sky Agent Network, an independent set of capital allocators who put USDS to work across strategies like onchain loans, Treasury bills, and lending liquidity.

What makes it measured rather than trust me? USDS is over-collateralized and backed by diversified collateral, so the backing is not a single bet.

The yield is risk-adjusted and governance-set, not whatever the market will pay this week.

Sky is the only stablecoin protocol with credit ratings from both S&P Global and Moody’s.

Both are sub-investment grade, S&P at B- and Moody’s at B3, which is the honest part: independent agencies see real credit and concentration risk, and they publish it.

The infrastructure also carries a decade-long track record with no smart-contract exploit, and governance is building a $150 million solvency reserve.

Adoption is following. In September 2026, Galaxy Digital, a Nasdaq-listed firm, put $100 million of sUSDS on its balance sheet and approved it as institutional collateral.

And yes, this ties back to AI. Sky’s own 2026 roadmap names AI-driven automation as a growth theme, alongside Sky Agents and tokenized credit.

The through-line is the same: use quantitative, verifiable risk work, not guesswork.

If you want to go beyond the savings rate, Sky Vaults and stUSDS sit further out on the risk curve, and Sky Ecosystem Rewards layers in token incentives.

One honest caveat for savers: the Sky Savings Rate is variable. It has moved from roughly 8.75% in early 2025 to around 3.6% in late September 2026, and governance can change it. Read the live number before you decide anything.

The agent frontier, and why it’s still risky

The next step is AI agents that do not just advise but act, holding permissions and moving funds on-chain, including around tokenized real-world assets. Promising, and early.

Every limit above gets sharper when the model can pull the trigger. An agent that can act can also be manipulated or prompt-injected. A misbehaving agent causes harm directly, not as a suggestion.

Safety research keeps landing on the same answer: strict, often deterministic guardrails and clear permissions, not unsupervised control.

Full autonomy, with your money, is not here, and it is not obviously what you want.

So, can AI actually manage on-chain risk?

Yes for the parts that need speed and scale. Not yet as a manager you can leave alone.

AI already simulates, monitors, and adjusts risk faster than any human team, and it is now core plumbing for major protocols.

But it works inside human-set bounds, it can be wrong or gamed, it struggles with the crises that matter most, and letting it act unsupervised introduces fresh risk.

The grown-up version of on-chain risk is not a black box making the calls. It is measurable risk, verifiable backing, and humans who own the guardrails, with AI doing the heavy lifting in between.

AI can manage parts of on-chain risk as augmentation inside human guardrails. It is not yet a risk manager you can walk away from.

Frequently asked questions

Q. Can AI manage DeFi risk?

A. It already manages important parts, like simulation, monitoring, and bounded parameter changes, but inside human-set guardrails, not as a fully autonomous manager.

Q. What does AI actually do in DeFi risk management?

A. It stress-tests parameters, watches markets in real time, auto-adjusts limits within approved bounds, rebalances curated vaults, and flags anomalies faster than people can.

Q. What are the limits of AI risk management in DeFi?

A. It only acts within human bounds, depends on data and oracles, can be gamed, struggles with novel tail events, and over-automation can itself spread risk.

Q. How does Sky approach risk?

A. Like other protocols, Sky relies on quantitative, increasingly AI-assisted risk modeling to set the parameters behind USDS and the Sky Savings Rate, and its risk is also measured externally with credit ratings from both S&P and Moody’s. That does not make it risk-free.

• • •

On-chain risk is getting more measurable every year. That is the real story, bigger than any single model. If you want to see what governance-set, verifiable yield looks like up close, you can put your stablecoins to work on sky.money and read the live numbers on the dashboard.

AI Risk Management in DeFi: What It Can and Can’t Do was originally published in Coinmonks on Medium, where people are continuing the conversation by highlighting and responding to this story.

I compared ETH funding across Kraken, Coinbase, OKX and Hyperliquid. The rates looked wildly different until I checked what each number was actually measuring.

Kraken was showing an ETH funding rate almost 39× higher than Coinbase CDE when I checked both within the same three-minute window.

Disclosure: AI assisted with research, drafting and editing. The market observations, screenshots and calculations in this article were collected by me on September 29, 2026.

At first, I thought I had found a huge discrepancy.

Then I put the rates on the same clock.

Coinbase CDE became 0.0064% over eight hours. OKX was 0.0100%. Hyperliquid was approximately 0.0104%.

Kraken was still 0.249376%.

That was when the comparison became more interesting.

The problem was not simply that four exchanges displayed different percentages. The contracts, funding intervals and even the meaning of the displayed numbers were different.

Before deciding that one rate was “high,” I had to work out what I was actually comparing.

Four screens, four numbers

Here is what I captured during the short observation window:

Kraken — ETH Perp
0.031172% / hr · hourly

Coinbase CDE — nano Ether perpetual-style futures
0.0008% · hourly

OKX — ETHUSD UM X-Perp
0.01% · 8 hours

Hyperliquid — ETH-USDC Perp
0.0013% · hourly payment cycle

At first glance, these look like four versions of the same metric.

They are not.

The first problem becomes obvious as soon as the funding periods are written next to the percentages.

Method note: I captured the four public venue interfaces between approximately 09:50 and 09:53 UTC on September 29, 2026. I did not place a trade. I recorded the displayed rates first, then checked each venue’s funding interval and contract documentation before normalizing the rates.

Why crypto funding rates are easy to compare incorrectly

A rate quoted for one hour cannot be compared directly with a rate quoted for eight hours.

The first mistake I made was comparing percentages before comparing their clocks.

Coinbase says its US perpetual-style futures funding rate is calculated hourly, based on the relationship between its futures and spot prices.

OKX states that funding for its X-Perps is settled every eight hours.

Hyperliquid uses another structure. Its documentation describes an eight-hour funding-rate formula, while funding is paid every hour at one-eighth of the computed rate.

That means the raw percentages need a common clock before they can be compared.

Using an eight-hour equivalent:

• Kraken: 0.249376%
• Coinbase CDE: 0.0064%
• OKX: 0.0100%
• Hyperliquid: approximately 0.0104%

Something interesting happens after normalization.

OKX and Hyperliquid, which initially looked different because of the way their rates were displayed, land very close to each other.

Kraken still does not.

And that led to a second problem.

Same ETH does not mean the same contract

The underlying asset was ETH in all four observations.

The derivatives were not identical.

Coinbase CDE’s product is a US perpetual-style futures contract with its own clearing and funding process. Coinbase says the funding rate is calculated hourly, while the resulting cash adjustments are processed through its clearing system.

The OKX screen I captured showed an X-Perp, a long-dated derivative designed to behave similarly to a perpetual contract while retaining a distant settlement date.

Hyperliquid uses its own perpetual structure, funding formula and margin system.

Kraken has its own contract specifications and funding mechanics as well.

So this sentence:

“ETH funding was X%.”

is incomplete.

The better questions are:

Which ETH contract? On which venue? Over what funding interval?

The same asset name does not make two derivative products economically identical.

The 39× gap was real

The cleanest comparison in my screenshots was Kraken versus Coinbase CDE because both were showing hourly funding rates.

Kraken:

0.031172% / hour

Coinbase CDE:

0.0008% / hour

Dividing one by the other:

0.031172 ÷ 0.0008 ≈ 38.97

So, in those observations, Kraken’s displayed hourly rate was roughly 39× the Coinbase CDE rate.

That calculation is real.

What it does not establish is that Kraken is always 39× more expensive, or that the two contracts should normally have the same funding rate.

The observation describes a moment.

It does not create a permanent relationship between the venues.

The 273% headline that the same screen undermined

The Kraken number becomes even more dramatic if I annualize it.

0.031172% × 24 × 365 ≈ 273%

That would make a very clickable headline.

It would also be extremely easy to misunderstand.

It does not mean a trader should expect to pay or receive 273% over the next year.

It means something much narrower:

If that single hourly rate somehow remained unchanged for an entire year, its simple annualized equivalent would be roughly 273%.

The same Kraken screen gave me a reason not to treat that annualized figure as a forecast.

Alongside the current funding rate of:

0.031172% / hr

the interface showed a next funding rate of 0.0064350% / hr.

The next displayed rate was roughly 79% lower.

So a single screen contained both:

• a current rate capable of producing an eye-catching annualized number;
• and a next displayed rate already far below it.

A mathematically correct annualization can still create an economically misleading impression.

The same problem appears when a trading strategy is tested on historical data: funding rates, borrowing costs and liquidation mechanics need to be treated as explicit execution assumptions rather than background noise. Forvest’s guide to backtesting vs forward testing explains why those assumptions can change what a strategy appears to achieve once it moves from historical simulation to live conditions.

I checked Kraken again one hour later

I also ended up with a small follow-up check.

I captured Kraken once at approximately 08:50 UTC and again at approximately 09:50 UTC.

In both public-interface screenshots, the displayed current funding rate was:

0.031172% / hr

and the next displayed rate was:

0.0064350% / hr.

That does not prove the rate normally remains unchanged for an hour.

Two observations are not a time series.

They only allow a much narrower statement:

In two Kraken screenshots taken roughly one hour apart, the same current and next ETH funding rates were displayed.

That distinction matters.

I can document what the interface showed twice.

I cannot use two screenshots to describe Kraken’s normal funding behavior.

Before sharing a funding screenshot, check three things

After doing this comparison, I would no longer start with the size of the percentage.

I would start with three checks.

1. What is the funding interval?

One hour?

Eight hours?

Something else?

A percentage without its funding interval is incomplete information.

Two rates can look dramatically different before normalization and much closer afterward.

2. What contract are you actually looking at?

“ETH funding” sounds like a single market-wide number.

It is not.

The venue, contract structure, settlement design, collateral and calculation methodology can all matter.

Even products designed to behave similarly can have different mechanics underneath them.

3. Is the number current, next or annualized?

Those labels answer different questions.

A current funding rate describes the rate associated with a particular funding period.

A next or predicted rate refers to another period.

An annualized rate is a mathematical transformation of a shorter-period rate.

It is not automatically an expected annual return.

And none of those numbers, by itself, predicts ETH’s next price move.

The useful number was not the biggest one

I started this exercise thinking the interesting result would be:

Which venue has the highest ETH funding rate?

That turned out to be the wrong question.

The more interesting finding was how easy it was to build a bad comparison from four legitimate screenshots.

Same underlying asset.

Similar-looking funding labels.

Different clocks.

Different derivative structures.

Different meanings.

The 39× gap was real between the two observed hourly rates.

But understanding what that gap meant required more work than dividing two percentages.

That is what I would check first the next time I see a funding-rate screenshot shared as evidence.

A funding-rate screenshot is not a comparison until you know what the number is measuring.

Educational analysis of public derivatives-market data. No trades were placed for this comparison.

FAQ

Are crypto funding rates directly comparable across exchanges?

Not always. Funding intervals, contract design, collateral, price references and settlement mechanics can differ. The interval should be normalized first, and the underlying contracts should be checked before drawing a conclusion.

Does a high ETH funding rate predict ETH’s next move?

Not by itself. Funding describes conditions inside a derivatives market for a particular period. It can provide useful context about positioning and contract pricing, but a single funding-rate observation is not a directional price forecast.

Crypto Funding Rates: I Found a 39× ETH Gap in Three Minutes was originally published in Coinmonks on Medium, where people are continuing the conversation by highlighting and responding to this story.

CoinGlass, Hyblock and CoinMarketCap describe different inputs. Here’s how to check the data before trusting a price-target claim.

Imagine a Bitcoin liquidation heatmap labeled 10:00, with a bright band at $90,000. At 10:15, a price chart reaches that level. Put the images side by side and the verdict seems obvious: the heatmap called it.

Disclosure: AI was used for research and drafting. Mobina Ebrahimi works in research and SEO at Forvest. Documentation checked September 28, 2026. The example above is hypothetical, not a market observation.

Two matching numbers are not yet a tested prediction. First, you need to know what produced the band, whether it was visible before the move, and what actually happened afterward.

The providers’ own documentation gives a useful starting point: similar-looking heatmaps can represent different evidence. That difference matters before you turn a colored zone into a trading decision.

Start with the data — not the color

The heatmaps discussed here map potential liquidation levels over time. They concern prices at which leveraged positions could be forcibly closed, not simply a list of liquidations that have already happened.

CoinGlass Model 1 starts with a calculation. Its aggregated heatmap documentation describes levels calculated from market data and liquidation leverage levels. The description does not establish that each band represents individually observed accounts. This is a statement about that endpoint, not every CoinGlass product.

Hyblock makes an estimation limit explicit. Its heatmap guide describes calculating levels and grouping them into price buckets. It warns that the tool predicts where levels open, but not where they close, and advises reading the displayed size relatively. A bright zone is therefore not a verified inventory of positions still open at that moment.

CoinMarketCap describes a position-based history. Its Bitcoin heatmap FAQ says it records liquidation levels from Hyperliquid positions on-chain roughly every five minutes. The page identifies Hyperliquid as its available coverage. It also says brightness is relative to the visible range, rather than an absolute dollar scale.

These are the providers’ descriptions, not an independent audit of their data. The comparison does not rank their accuracy. It establishes why “liquidation heatmap” is not the name of one universal dataset.

A model is not automatically useless. A position-based chart is not automatically complete. Neither becomes a forecast because its brightest band looks convincing.

A real position can still have a changing liquidation level

Observed positions answer a question about the input. They do not remove the conditions governing liquidation.

Hyperliquid’s documentation says liquidation becomes possible when account equity falls below maintenance margin — the minimum required to keep positions open. It uses a mark price, a reference price that can differ from the latest trade on the order book.

The same document explains that funding payments, or changing gains and losses on other positions in a cross-margin account, can alter the liquidation price. In cross margin, positions share collateral.

That matters when someone points to a candle touching a heatmap band. Was it the right venue and the right price reference? Did the relevant positions and account conditions still exist?

CoinMarketCap’s FAQ also distinguishes liquidation levels from orders already resting on an order book. A mapped level is not an executed liquidation. Reading it from a real position does not settle when — or whether — that event will happen.

The timestamp test: what was visible before the move?

Return to the hypothetical $90,000 band. The two screenshots invite a conclusion, but three checks separate a persuasive picture from supporting evidence.

Before the move: establish when the heatmap was captured, what data it contained, and which settings were selected. A historical chart generated later can show earlier times without proving what a reader could see at those times. That is a difference in evidence, not an accusation that a provider changed its history.

At the level: separate a price touch from a liquidation event. A claim that positions were liquidated needs an event record for the relevant venue and time. Two pictures displaying $90,000 cannot supply that record by themselves.

Across attempts: a single apparent success does not establish a dependable rule. Define the target, deadline and selection rule before checking the outcome, then include the cases that missed — not just the screenshot that worked.

A contemporaneous screenshot can support what was displayed. An event record can support what happened. Testing a forecast requires a defined rule and more than one selected outcome. Those are different jobs.

Give the screenshot a three-line record

You do not need to build a trading model to ask better questions. Before saving or sharing a heatmap, add three short notes:

Source: provider, exact chart or model, asset and venues. Is the input estimated or position-derived? One venue’s chart is not the whole Bitcoin market.

Time: capture time and time zone, latest data time if available, lookback, filters and legend units. Was this captured then, or generated later? Leave missing details unknown.

Claim: what are you saying the image demonstrates — a concentration of mapped levels, an actual liquidation, or a prediction? Name the extra evidence needed to support that claim.

For the opening example, even a genuine 10:00 capture might support only this: “The chart displayed a band near $90,000 before the later price touch.” That is a narrower — and more defensible — statement than “This chart reliably predicts Bitcoin.”

The same distinction between an original source and somebody’s interpretation runs through Forvest’s guide to analyzing crypto news. Here, the caption attached to a chart deserves as much scrutiny as the chart itself.

A useful map does not have to be a forecast

A heatmap can help locate concentrations within its stated method and coverage. It gives you a specific question to investigate. That is useful without pretending the display tells you where to buy or sell.

The bright band in our example has not become more or less colorful. What has changed is the claim we can responsibly make about it: first identify the input, then establish the timing, then check the outcome.

Knowing where a position could be liquidated is not the same as knowing where Bitcoin will go.

Educational analysis of data interpretation, not a recommendation to open, close or size a trade. Leveraged trading can result in substantial losses.

Bitcoin Liquidation Heatmaps: What the Bright Bands Really Show was originally published in Coinmonks on Medium, where people are continuing the conversation by highlighting and responding to this story.