Sui Foundation said two separate bugs introduced in the v1.72 upgrade were behind three mainnet halts that disrupted network operations on Thursday and Friday. The foundation said no user funds were ever at risk, and no committed transactions were reversed during recovery.
Gas-charging flaw triggered the first two outages
The first halt started at about 7 a.m. PT on Thursday and lasted until roughly 1:30 p.m. PT. According to the foundation, the issue involved gas charging and the newly added address balances feature. When transactions competing for the same funds hit an InsufficientFundsForWithdraw error, the transaction itself was canceled, but later gas-processing steps still attempted to spend that same balance.
That created a negative balance during settlement. Validators then crashed, which brought the network down. To restore service quickly, the Sui Core Team proposed an interim patch on Thursday, though it said a known low-probability risk remained in that temporary fix.
The network hit a variation of the same issue on Friday morning, causing the second halt before validators rolled out a more complete fix.
A separate randomness-state bug caused the third halt
After the second recovery, the network ran normally until the next scheduled epoch transition on Friday afternoon. At that point, a different bug triggered the third outage. Sui Foundation said validators restarted to install the Friday fix but did not meet participation requirements for distributed key generation, so randomness was disabled for that epoch as intended.
A latent bug then prevented validators from preserving that status across restarts. Transactions that depended on randomness could not be resolved, the queue built up, and validators were unable to close the epoch. This outage lasted from about 1:30 p.m. PT until 7:20 p.m. PT on Friday.
Fixes applied and a new forced-close safeguard added
Validators later fixed both the gas-charging issue and the randomness-state bug. Sui Foundation also added a mechanism that lets validators force-close a stalled epoch when needed.
The report pointed to three areas that need work: gas-charging logic, resilience at epoch end, and failure containment. It also said AI agents helped engineers diagnose the incidents faster by analyzing validator logs, production data, and network metrics during the outages.