Ops Monitor

The bridge used a single verifier node to check cross-chain messages before releasing funds. Attackers compromised the RPC endpoints that verifier relied on, fed it a message with no matching burn behind it, and drew 116,500 rsETH out of escrow, roughly eighteen per cent of supply. No contract had a bug.

RPC integrity is a named surface here. Independent providers are asked for the same block hash and the same key state reads, and divergence or staleness is itself a critical finding. A verifier reading from endpoints that disagree with the network is the alert, hours before a forged message clears.

It takes more than one wall. An Asset Ratings floor caps a one-of-one verifier outright, whatever else the asset scores, so this configuration is refused at integration rather than discovered at the loss.

A Safe{Wallet} developer's workstation was compromised and tampered JavaScript was uploaded to the bucket serving the wallet interface. Seventeen days later the signers approved what that interface showed them, an ordinary transfer, while the payload sent to their hardware devices carried a delegatecall that replaced the cold wallet's implementation.

Two independent checks, either enough. The served bundle changed and its hash no longer matches what was reviewed. And a delegatecall from a Safe is unconditionally critical here, with no model consulted and no threshold to tune, because it is never benign.

macOS malware on contributor machines displayed legitimate transaction data in the Safe front end while sending different payloads to the hardware wallets for signature. The procedure was followed exactly, on hardware, by people doing their jobs correctly. Malicious contracts were staged on four chains; two executed.

A per-signer baseline scores time of day, inter-signature latency, signing velocity, submitting address, gas habit and target novelty, so a coerced or spoofed approval breaks a pattern even when the signature is perfectly valid. The ownership transfer it authorised is separately an unconditional critical.

A proposal presented itself as identical to one the DAO had already approved, but carried an extra function. After the vote passed, the proposal contract was destroyed and redeployed at the same address with different code, granting the attacker 1.2 million votes and complete control of governance.

A proposal is judged by what it executes on forked state before the vote closes, not by what its description claims. A payload whose deployed bytecode no longer matches what was reviewed is a critical finding on its own, and the simulated diff reaches voters while voting is still open.

A compromised Cloudflare API key let an attacker inject a script into the application's own front end. The script quietly added an unlimited spend approval for the attacker's address to transactions users were already making. Around 500 wallets signed one.

The served JavaScript changed and its hash stopped matching the reviewed bundle, which is a critical finding on the front end your users actually load, alongside DNS records and certificate transparency logs. The class is covered chain-agnostically, because a domain does not care which virtual machine you run.

It takes more than one wall. Live Detection's wallet lane closes the other half: an unlimited approval to a spender first seen minutes ago is blocked at the signing screen, even when the site is already compromised.