Four of tonight's stories are about trusting an authority you inherited
Victoria Shinder2 min readFour unrelated items tonight, and each one turns on something being treated as authoritative because of where it came from rather than because anyone checked it.
Bitget lost $387.5 million and no private key was stolen. The attackers forged internal transfer requests convincingly enough that the approval path treated them as routine; the signing infrastructure then did exactly what it was asked. The trusted input was the exchange's own request traffic. Every control downstream of that - multisig, hardware custody, withdrawal limits - operates on the assumption that the request in front of it is real, and none of them verify it independently.
Solana's SIMD-0376 is the same problem in a longer time frame. Consensus behaviour was inherited from ed25519-dalek's verify_strict - not a specification, just the code everyone ran. The amendment replaces it with a cited equation and explicit rejection criteria, and in doing so finds something neither document contained: adopting ZIP-215 verbatim would have made Pubkey::default(), the System Program ID, a signable public key. Two correct specifications, one broken seam, visible only to someone holding both.
XRPL's #636 dissolves a standalone closed-ended vault draft into patches on XLS-65 and XLS-66, deleting 620 lines and adding 453. The draft had to restate base vault behaviour in order to describe its variations, and a restatement is an authority that looks official and drifts. Citing the parent instead removes the second description before anyone implements against it.
And the FCA closed twenty-one CFD firms for the retail version of the same thing. These were firms doing little UK business whose authorisation existed mainly to make linked overseas companies look supervised. The customer inherited trust from a brand and a badge, and the regulator's remedy was to make the badge unusable rather than to fine anyone.
The common failure is not carelessness. In each case the inherited authority was genuinely authoritative somewhere - a real request format, a well-reviewed library, a correct base spec, an actual UK licence. What was missing is the step that asks whether it is authoritative here, for this decision. That question is cheap to ask and almost never scheduled, because the artefact arrives already looking finished.
Two practical readings. For anyone operating custody: name, explicitly, what verifies that an internal transfer request is genuine before a signer sees it - and if the answer is the system that produced it, Bitget's attack works on you. For anyone importing a standard: write down which of your own conventions the standard does not know about. Solana found its System Program ID that way, in a document review, which is the cheapest place such a finding can possibly turn up.