Apple — the privacy-as-product company of the #077 FBI standoff — announced it would scan photos ON-DEVICE for child-sexual-abuse material before iCloud upload, using perceptual hashes matched against a known-CSAM database, with human review past a threshold. The backlash from the security community was immediate, near-unanimous, and NOT about the target (nobody defends CSAM): it’s about the ARCHITECTURE — client-side scanning builds the searching machinery INTO the device, and the machinery cannot know what database it matches against (today CSAM hashes from NCMEC; tomorrow, any government’s “terrorist content” list, with compliance leverage Apple has already demonstrated bending to in China). The #077 principle returns inverted: Apple argued then that a capability built for good guys can’t be contained; critics now quote Apple’s own brief back at it. The file’s read: the CRYPTOGRAPHIC design is genuinely clever (threshold secret-sharing, safety vouchers — the engineers did careful work); the POLICY architecture is the vulnerability, and “we would refuse such demands” is a promise, not a mechanism (#196’s AUP lesson: capabilities outlive intentions; the sword you forge is the sword that exists). Apple has already paused the rollout under pressure. The precedent debate — where does content-matching belong: device, transit, or cloud — is now permanently open, and every E2E messaging platform’s future runs through it.
The fortnight’s crypto absurdity, filed for range: the Poly Network hack — $600M+ drained via a cross-chain contract flaw, the largest DeFi theft ever — resolved with the attacker RETURNING essentially everything, conducting an interview-by-embedded-transaction-notes, claiming it was “for fun” and to expose the bug, and being offered a $500k bounty and a job title (“Chief Security Advisor”) by the protocol. Code-is-law jurisprudence currently operates at the maturity level of a pirate ship electing its quartermaster (#110’s forever-bugs file: immutable code, negotiable everything else).
TIL: perceptual hashing (NeuralHash) collision mechanics — within DAYS of the announcement, researchers produced synthetic collisions against the extracted model. Any classifier deployed at civilizational scale becomes an adversarial target at civilizational scale (#145’s confusion-matrix politics; #079’s Tay clause — your users are your red team, forever).