Term Finance had a seven-day timelock and an LP veto. The wrapper around them became the attack surface.
Two security firms put the drain at roughly $8.5 million. Term Labs has confirmed a governance exploit but not the amount, not which role was abused, and not why a week-long delay and a depositor veto did not stop it. Here is what is established, what is estimated, and what nobody has said yet.
A protocol that documented a seven-day delay on governance actions and gave its liquidity providers the right to veto queued transactions lost control of its vaults anyway. That combination — strong controls on paper, a successful governance exploit in practice — is the reason this incident is worth more than its dollar figure.
What is confirmed. Term Labs acknowledged the incident publicly, stating that it was aware of a governance exploit impacting Term vaults and that further detail would follow an investigation. It did not name the affected vaults and did not confirm a loss figure. The Block reported the exploit on Sunday 23 August, describing an unauthorised movement of ether and stablecoins out of the protocol's Strategy Vaults.
What is estimated, and by whom. PeckShield traced roughly 2,843 ETH — about $6.9 million at the time — plus 1.68 million USDC, subsequently converted to DAI. CertiK put the total at approximately $8.5 million. Those are two independent onchain estimates, not an audited or protocol-confirmed loss, and the distinction matters for anyone modelling recovery: onchain tracing measures what left an address, not what is ultimately unrecoverable, and it can revise. The Block noted prior total value locked of roughly $12.45 million across chains, about $8.8 million of it on Ethereum, which frames the drain as a large share of a small protocol rather than a systemic event.
What was hit, and what was not. The affected contracts were Term's Strategy Vaults, ERC-4626 tokenised vaults built on Yearn V3 infrastructure. Yearn stated that the exploit occurred via a custom governance wrapper, distinct from standard Yearn operations. This is the single most misread fact in an incident of this shape: shared vault infrastructure was underneath, but the exploited surface was the bespoke permission layer above it. Nothing in the record supports a claim that standard Yearn vaults were compromised, and nothing here should be read as one.
The controls that existed. Term's own developer documentation describes a deliberately layered design. Two roles are separated: a manager, which executes auction operations and manages offers, and a governor, which oversees governance actions, risk-parameter configuration and emergency controls — including setting maturity thresholds, reserve ratios, concentration limits, collateral requirements and token blacklists. Governance actions do not execute directly. The documented path runs from a Proposer Safe, through a Zodiac Delay Module carrying a seven-day timelock, to a Governor Safe, and only then to the vault. During the delay, vault liquidity providers sit as DAO members with the documented ability to review a queued proposal and, by vote, skip it — invalidating the transaction in the Delay Module before it can execute.
On paper that is a strong design. A week is a long time to notice a hostile parameter change, and a veto held by the people whose money is at stake is the right party to hold it.
What nobody has said. Term has not disclosed which role the attacker obtained, whether the seven-day delay was in the path of the transactions that drained the vaults, or why the LP veto did not fire. Each possibility implies a different lesson, and they are not interchangeable. If the timelock was in the path and simply expired unchallenged, the failure was one of monitoring and participation — the control worked and nobody used it. If the drain travelled a path the Delay Module did not govern, the failure was one of scope: a control that does not cover every privileged route is not a control over the vault, only over part of it. If a role was transferred or acquired in a way that made the attacker legitimate to the system, then the timelock and the veto were operating exactly as designed, on behalf of the wrong principal. Until Term publishes a post-mortem naming the path, treating any one of these as established would be a guess.
PeckShield reported that the funds moved to a single address initially seeded with 2 ETH withdrawn from Tornado Cash, which tells you the operation was cheap to fund and prepared for attribution resistance. It says nothing about which of the three failure modes above occurred.
The control-design lesson. The generalisable point does not depend on resolving that question, which is why it is worth stating now. A delay is not a defence; it is an opportunity to defend, and it converts into a defence only when someone is watching the queue and is both able and motivated to act inside the window. A veto distributed across liquidity providers has a participation requirement that a smart contract cannot enforce: dispersed depositors in a protocol of this size are not a standing security team. And a governance wrapper that assigns roles over a vault is part of that vault's trust boundary even when the vault itself is well reviewed — composition moves the attack surface to whichever layer received the least scrutiny, and a bespoke permission layer written for one protocol is almost always that layer.
The uncomfortable corollary for readers assessing any curated-vault product: the audit status of the underlying vault standard is not the security of the deployment. The question to ask is which contract can change the vault's parameters or move its assets, what delays that path, and who in practice is watching during the delay.
Watch points. First, a Term Labs post-mortem that names the compromised role and states whether the timelock was in the exploited path — that single disclosure separates the three failure modes. Second, whether the ETH and the roughly 1.68 million DAI move again, and whether any recovery or negotiation follows, which would change the estimated figures. Third, whether the two security-firm estimates converge or are revised once the full transaction set is reconstructed. Fourth, whether other protocols running custom governance wrappers over shared vault infrastructure publish reviews of their own privileged paths — that is the response that would make this incident worth its cost to the wider market.