Risk in DeFi: August post-mortem

---

Overview

This review examines eleven of August 2026’s most notable onchain incidents in date order, measuring each against Credora’s risk assessment methodology to reach a verdict on whether the methodology would cover it. It serves two ends: an account of how and why each failure happened, and a standing test of the methodology against realized losses. The verdicts that fall short matter as much as the ones that land, since they mark where the framework must evolve. A scorecard follows at the end.

August set a record. By PeckShield’s count, the month saw 50 major incidents, the highest of any month in 2026, and roughly two-thirds more than July, even as total losses fell about half, to around $136M. The average loss dropped from roughly $9M in July to about $2.7M, indicative of many small hits as opposed to a few large incidents. Trackers diverged on the month’s total, with CertiK reporting a higher $215M across all of crypto and about $144.6M in DeFi.

---

Shifting attack vectors

One feature defined August: how many chains halted themselves. Cronos, Fogo, Realio, Injective, BounceBit, and the three Cosmos EVM chains all stopped producing blocks during the month. A small or curated validator set can freeze a network within the hour, which capped several losses; Tectonic’s attacker escaped with about $6M of a $74M exploit. The same design concentrates control, and in Cronos’s case restarting meant rolling the chain back, discarding close to 11,000 blocks of unrelated users’ transactions along with the attacker’s.

---

1. Neutrl (NUSD / sNUSD), reserves paused, August 13

Neutrl, a synthetic-dollar protocol issuing NUSD and its staked form sNUSD, abruptly paused minting and redemptions on August 13, citing circumstances affecting protocol reserves. sNUSD fell sharply on the news, and the protocol’s total value locked, already down from a February peak near $200M to about $53M, was frozen. In a more detailed statement two weeks later the team said roughly $27M in liquid assets remained, with the rest in a strategy position that was “not presently liquid,” and that it could not confirm the timing, amount, or recovery value.

The timing and the disclosure drew scrutiny. A suspected team address had withdrawn liquidity from the NUSD-USDC pool on Curve shortly before the freeze, and an affected depositor noted that the book had shown about $137M in June against $27M liquid now, asking when it had become undercollateralized. Neutrl called the event an illiquid strategy position rather than a hack, and offered an early-redemption mechanism targeted for early September. The failure was a reserve that could not be redeemed on demand, and a transparency gap about what stood behind the token.

Methodology mapping

→ Reserve Transparency & Management: the dimension scores whether a token’s backing is verifiable onchain, only vouched for by third-party attestation, or effectively unverifiable, and whether reserves are liquid enough to honor redemptions. Neutrl’s reserves were attested rather than provable onchain, and when the strategy turned illiquid it disclosed neither the affected asset nor the size of the loss.

→ Redemption & Liquidity: the framework scores whether holders can exit at par and how deep secondary liquidity is. A synthetic dollar that freezes redemptions while its book turns illiquid, with liquidity thinning on its main pool, is the redemption and thin-liquidity profile these modifiers capture.

Verdict: COVERS. A synthetic dollar backed by an illiquid strategy, unredeemable on demand and thin on disclosure, is the reserve, redemption, and liquidity profile the framework scores; June’s msUSD depeg carried the same lesson. This was not hypothetical for Credora: our internal assessment, as of June 24, had rated NUSD a C+ on the A+ to D scale, assigning high risk to the reserve assets and strategies behind the token, flagging that reserves were third-party assured only rather than verifiable onchain, and noting that redemptions were already queued with occasional delays before the pause.

---

2. Term Finance (Term Labs), est. ~$8.5M, August 23

Term Finance, an Ethereum fixed-rate lending protocol, lost about $8.5M when an attacker captured governance of its Meta Vaults and voted to redirect vault funds to an address they controlled. The attacker drained roughly 2,843 ETH (about $6.87M) and 1.68M USDC, swapped to DAI. Term Labs then permanently shut Meta Vault deposits and revoked DAO governance roles, keeping withdrawals open.

The mechanism was governance, not code. On August 17, six days before the drain, the attacker bought about 0.4852 tmvETH for roughly 0.5 ETH (around $951) and staked it, enough to control approximately 90.66% of the votes in the vault’s governance pool. The vault had safeguards, a seven-day execution delay and an LP veto, but participation was so thin that no one exercised the veto, and once the delay elapsed the proposal executed.

Methodology mapping

→ Governance (Capture Risk): the modifier penalizes governance that can be captured cheaply relative to the value it controls, and weights the penalty higher where participation is low. A roughly $951 stake commanding 90.66% of the vote over an $8.5M vault is that exposure directly, and the low participation the metric weights is what left the seven-day delay and LP veto unused.

Verdict: COVERS. The framework scores governance capture and weights it up where participation is thin, which is what left Term’s safeguards unused: a delay and a veto only work if someone acts on them. July’s BonkDAO was partially covered because the execution timelock itself was unscored; Term had that timelock and lost anyway, which is why the metric weights participation rather than the presence of a delay.

---

3. PT-reUSD liquidation cascade (Morpho / Pendle), est. ~$36.4M liquidated, August 25

A single trader spent about $320,000 and set off roughly $36.4M in liquidations. Between 04:28 and 04:37 UTC on August 25, wallet 0x854e…690d ran 11 consecutive Pendle trades, converting around $320,000 of SY-reUSD into more than 9.5M YT-reUSD, which pushed the market’s implied yield past 20% and cut the price of PT-reUSD by about 3%. That was enough: 33 Morpho liquidation events followed, repaying about $36.14M of debt (roughly $35.19M USDC and $960K USDT) and seizing over 38M PT-reUSD as collateral.

The damage came from leverage meeting a movable price. Morpho markets using PT-reUSD as collateral carried a high liquidation LTV, and loopers had borrowed right up to it, leaving positions a buffer of under 10%. The PT oracle followed a 15-minute market average, so a roughly 3% dislocation pushed dozens of leveraged positions across the liquidation threshold at once. Morpho’s core contracts were not exploited and reUSD did not depeg, and Pendle called the event market dynamics and high leverage rather than an oracle bug. That the oracle worked as designed is no defense: the 15-minute market-following price a $320K trade could move, the liquidation LTV set high enough for loopers to stack to a sub-10% buffer, and the onboarding of a thin-liquidity principal token as collateral were all risk-parameter choices, and each failed.

Methodology mapping

→ DeFi Composability & Reflexivity: this is the levered-collateral spiral the dimension scores. Positions looped to the edge of a high liquidation LTV on a single principal token turned a 3% price move into $36M of forced liquidations.

→ Collateralization: the gap between a position’s LTV and the market’s liquidation LTV is the buffer this dimension reads, and here it was thin on a token whose price a modest trade could move. A high liquidation LTV on a volatile principal token, with loopers stacked beneath it, is an aggressive configuration well before any trade tests it.

Verdict: COVERS. Leverage and collateral parameters are scored, so positions looped to the edge of a high liquidation LTV on a principal token would have rated poorly in advance. Morpho’s core code did not fail and the oracle did not malfunction, but the choices around them did: the manipulable price window, the aggressive liquidation LTV, and the thin-liquidity collateral are all scored, and were visible before the trade that triggered them.

---

4. Moonwell (MAMO), est. ~$8.7M, August 27

Tectonic, the largest lending protocol on Cronos, was hit for roughly $74M, the fourth-largest crypto theft of 2026, when an attacker inflated the price of the TONIC token about a hundredfold and borrowed real assets against it. Cronos halted the entire chain; the attacker had bridged only about $6M to Ethereum before the freeze, leaving roughly $68M stranded, and Cronos restarted by rolling state back to before the attack. Tectonic’s deposits, about $121.7M before the incident, fell close to 97% within a month.

Over about 20 minutes the attacker pumped TONIC roughly 100x, then posted it as collateral to borrow genuine assets. TONIC was Tectonic’s own governance token, carried at a 20% collateral factor on very thin liquidity, so the manipulation was cheap relative to what it unlocked, the same thin-liquidity collateral play seen at Moonwell days earlier, at far larger scale. Cronos’s response, a full-chain halt and a state rollback that discarded close to 11,000 blocks, contained the loss but reversed unrelated users’ transactions along with the attacker’s.

Methodology mapping

→ Oracle Risk & Collateralization: pricing collateral from a manipulable, thin-liquidity token, and accepting a protocol’s own governance token as collateral at a meaningful factor, are the collateral-valuation failures these dimensions score. The protocol lent against a price the market did not support, as Moonwell did.

→ Circuit Breaker: an outflow cap or a borrow limit tied to a price-sanity check is the guard this dimension scores. None constrained the borrow, and the only brake was a chain-wide halt after the fact.

Verdict: COVERS. Collateral valuation from a manipulable market, a self-token accepted as collateral, and the absence of an onchain circuit breaker are all scored; a lending market taking a hundredfold-inflated token as collateral would have rated poorly before August 30. The chain halt and rollback were platform-level incident response, not a control the protocol held.

---

5. Tectonic (Cronos), est. ~$74M, August 30

Tectonic, the largest lending protocol on Cronos, was hit for roughly $74M, the fourth-largest crypto theft of 2026, when an attacker inflated the price of the TONIC token about a hundredfold and borrowed real assets against it. Cronos halted the entire chain; the attacker had bridged only about $6M to Ethereum before the freeze, leaving roughly $68M stranded, and Cronos restarted by rolling the state back to before the attack. Tectonic’s deposits, about $121.7M before the incident, collapsed by almost 97% within a month.

Over about 20 minutes the attacker pumped TONIC roughly 100x, then posted it as collateral to borrow genuine assets. TONIC was Tectonic’s own governance token, carried at a 20% collateral factor on very thin liquidity, so the manipulation was cheap relative to what it unlocked. In essence, it was the same thin-liquidity collateral play seen at Moonwell days earlier, but at a far larger scale. Cronos’s response, a full-chain halt followed by a state rollback that discarded close to 11,000 blocks, contained the loss but reversed unrelated users’ transactions along with the attacker’s.

Methodology mapping

→ Oracle Risk & Collateralization: pricing collateral from a manipulable, thin-liquidity token and accepting a protocol’s own governance token as collateral at a meaningful factor are the collateral-valuation failures these dimensions score. The protocol lent against a price the market did not support, exactly as Moonwell did.

→ Circuit Breaker: an outflow cap or a borrow limit tied to a price-sanity check is the guard this dimension scores. None constrained the borrow, and the only brake was a chain-wide halt after the fact, a blunt instrument that also reversed innocent activity.

Verdict: COVERS. Collateral valuation from a manipulable market and the absence of an on-chain circuit breaker are both scored, so a lending market accepting a hundredfold-inflated token would have rated poorly well before August 30. The chain-level halt and rollback were incident response at the platform layer, not a control the protocol itself held; the standing weakness was assessable in advance.

---

6. Realio Network (RIO), est. ~124.4M RIO, August 25

Realio Network, a layer-1 built for tokenized real-world assets, disclosed on August 25 that its legacy realio.fund web platform had been attacked. About 124.4M RIO moved out of Realio-controlled wallets and some user accounts across multiple chains, and the foundation treasury was hit. Realio halted its native chain, closed its bridges, and warned users off the RIO deployments on Algorand and Stellar. RIO fell about 18% over the following week, though the team said the attacker sold little before balances were frozen.

The vector was the legacy platform, and the loss was private keys. Realio stated that the attacker gained access to the private keys of user wallets and to the keys controlling the $RIO bridge reserves on Stellar and Algorand, through the deprecated realio.fund web app, not through the chain itself. Validators halted the network as a precaution, and the Algorand and Stellar deployments will be deprecated, with those tokens never re-entering circulating supply. The affected surface was the platform’s custody of keys, not the L1’s consensus.

Methodology mapping

→ Key Custody Environment: the metric scores how an issuer’s operational keys are held and how far a single compromise can reach. Keys to user wallets and to bridge reserves, both reachable through one legacy platform, are the custody-concentration failure this dimension assesses, the same class as June’s Humanity Protocol, where a multisig on one laptop undid the appearance of distributed control.

→ Circuit Breaker: the chain halt and bridge closures contained the spread after the fact, but nothing capped movement out of the compromised wallets and reserves before validators intervened.

Verdict: COVERS. Custody of an issuer’s operational keys, and how far a single compromise reaches, are scored operational dimensions; keys to both user wallets and bridge reserves sitting behind one legacy platform would have rated poorly in advance. This was a theft of the keys themselves, not a front-end deception, which keeps it inside the rating. Because Realio is an RWA network the custody question is central: the value at stake is the tokenized asset base itself.

Summary

Incident	Date	Estimated loss	Root cause	Primary Credora dimensions	Verdict
Neutrl (NUSD / sNUSD)	Aug 13	No loss disclosed; mint/redeem paused	Reserves tied to illiquid strategy; liquidity pulled before freeze	Reserve Transparency, Redemption, Liquidity	Covers
Term Finance	Aug 23	~$8.5M	~$951 stake bought 90.66% of a vault vote; delay and veto went unused	Governance (capture)	Covers
PT-reUSD cascade (Morpho / Pendle)	Aug 25	~$36.4M liquidated	$320K Pendle trade moved PT ~3%; loops near a high LLTV cascaded	Composability & Reflexivity, Collateralization	Covers
Moonwell	Aug 27	~$8.7M	MAMO pumped ~40x; borrowed against inflated collateral on Base	Oracle Risk, Collateralization	Covers
Tectonic (Cronos)	Aug 30	~$74M	TONIC pumped ~100x; borrowed against it; chain halted and rolled back	Oracle Risk, Collateralization, Circuit Breaker	Covers
Realio Network	Aug 25	~124.4M RIO	Legacy webapp compromise; user-wallet and bridge-reserve keys stolen	Key Custody Environment	Covers
Other events…
Injective	Aug 31	~$4.9M	Market-ID collision from delimiter-free concatenation in settlement logic	Smart Contract Risk	Covers
Fogo Foundation	Aug 28	~$2.9M (400M FOGO)	Foundation key compromise moved ~10% of supply; mainnet halted	Key Custody Environment	Covers
Full Sail / Switchboard	Aug 29	Undisclosed	Switchboard oracle feed compromise; feeds shut on 4 Move chains	Oracle Risk	Covers
BounceBit	Aug 19-20	~$3M (286.5M BB)	Evmos chain-stack authorization bug; chain retired	None (chain framework, unscored)	Out of scope
Cosmos EVM (MANTRA, TAC, KiiChain)	Aug 20-22	~$11M+ combined	Shared cosmos/evm module flaw; vesting-account balance underflow	None (chain framework, unscored)	Out of scope

The six above are covered in full; the five under Other events are summarized here. Injective was a market-ID collision from delimiter-free concatenation in its own binary-options settlement logic, the same class as July’s Wanchain and, as the protocol’s own code, covered smart-contract risk. Fogo and Full Sail repeat patterns already covered in full: a foundation key that moved about 10% of supply (Key Custody Environment), and a compromised Switchboard oracle feed that a dependent vault inherited (Oracle Risk). BounceBit and the Cosmos EVM cluster are out of scope, both chain-framework bugs discussed below.

---

How to read the verdicts

Covers means the failure maps to a risk the framework scores, and a poor score would have been assignable before the event. Nine of the eleven qualify. Coverage does not mean the rating times the exact transaction, names the defective line, or predicts when a key is phished; those belong to execution monitoring, code audit, and forecasting. It means the standing weakness was scorable in advance.

Partially covers means the framework sees part of the risk but not the dominant part. No August incident fell here. July’s BonkDAO was the recent example: the governance-capture side was scored, but the execution timelock that would have made the drain contestable was not. Term Finance is the contrast this month; it had the timelock BonkDAO lacked, so its failure lands on the capture and participation factors the metric already scores, and it is a full Covers.

Not covered is the most consequential verdict: an incident within the framework’s remit that it failed to flag at all, a false negative rather than a boundary. No August incident fell here. The month’s chain-framework bugs were weighed against this verdict and placed out of scope instead, because their failures sat in reused platform code beneath the asset, not in a dependency the rating tracks such as an oracle or bridge.

Out of scope means the failure sits on a layer an asset rating does not read. Two August incidents fell here, both chain-framework bugs: BounceBit’s Evmos authorization flaw and the shared Cosmos EVM module bug behind MANTRA, TAC, and KiiChain. Injective marks the line: a bug in a protocol’s own logic, even an L1’s own exchange module, is scored smart-contract risk and is covered, while a defect in the third-party stack a chain runs on is not. The month also produced two patched advisories with no rateable onchain loss, a Ledger hardware-wallet firmware flaw and a Bitcoin Lightning node flaw, both beneath what an onchain asset measures, and the Coinsbuy breach, a centralized-exchange custodial theft rather than an onchain asset failure.

---

Key takeaways

August’s record was in frequency, not severity: 50 incidents, the most of any month in 2026, averaging about $2.7M each against roughly $9M in July. The dollar total was concentrated, with Tectonic, the PT-reUSD cascade, and Moonwell accounting for roughly $119M, all three driven by moving a price a lending market trusts for collateral. Nine of the eleven cases mapped to risks Credora prices in advance: collateral valuation and reflexivity (Tectonic, PT-reUSD, Moonwell), key and platform custody (Realio, Fogo), oracle-provider dependency (Full Sail), governance capture (Term Finance), reserve and redemption failure (Neutrl), and a settlement-logic bug (Injective). None is new; all repeat patterns the June and July editions mapped.

The two that did not map share a layer. BounceBit and the Cosmos EVM cluster were chain-framework failures, defects in the Evmos and cosmos/evm stacks the affected projects inherited rather than wrote, out of scope for the same reason July’s Coldcard was: they sit beneath the asset a rating reads. Injective is the contrast, a superficially similar delimiter-free identifier bug that lived in Injective’s own exchange logic, and so covered smart-contract risk. Where July surfaced a genuine refinement in BonkDAO’s unscored timelock, August surfaced no blind spot; its off-map cases were boundary questions. Whether chain-framework risk deserves its own platform-level overlay, separate from the asset rating, is a fair question the month raises, but it is not a gap in the asset methodology.

Beneath the featured events sat a long tail the framework would also read but that was too small to carry a section: an Ajna liquidation-accounting manipulation, an aquifer wallet compromise on Solana, and flash-loan spot-price drains on legacy Arrakis and Visor/Gamma vaults, most under a couple of million dollars, each a variant of trusting an instantaneous price or a single key.

Frequently Asked Questions

How does Credora assess risk in DeFi beyond the asset layer?

Credora’s methodology accounts for both onchain and offchain risk surfaces, including such dimensions as Key Custody Environment, Reserve Transparency, Counterparty Risk, and Peg Track Record. As indicated by the Coldcard exploit, some further layers of the overall blockchain ecosystem remain beyond the methodology’s scope, at least for now. The attack serves as a regrettable, but useful reminder to avoid concentrating your capital in a single wallet, even if it’s a cold one.

How does the Probability of Significant Loss (PSL) measure risk in DeFi?

The PSL is Credora’s main risk metric for markets and vaults, assessing the annualized probability that a position loses more than 1% of principal over a one-year horizon. A companion metric, the Probability of Default (PD), applies the same logic to individual assets. Both convert qualitative risk factors, from the quality of smart contract audits and key custody environment to reserve transparency, peg performance, and counterparty dependencies, into a single quantitative value that can be used to compare different markets, vaults, and assets. The result is converted into a risk rating, which runs from A+ to D.

What were the main DeFi exploit categories in July 2026?

While July’s single biggest exploit stems from a firmware vulnerability in a hardware cold wallet, it’s difficult to be qualified as a DeFi attack since the vulnerability was in a layer below the smart contracts powering decentralized applications. As far as “DeFi proper” goes, bridge exploits remained the most expensive category, while wallet and key compromises overtook smart-contract exploits in terms of single attack cost.