In 2016 the FBI took Apple to court over a locked iPhone and lost interest before it lost the case. An outside firm sold the bureau a way into the device, the order was withdrawn, and the largest encryption fight in a decade ended without anyone deciding anything. Matthew Green argues this week that the arrangement which made that possible — a reliable commercial supply of working exploits — is about to run out.
The forecast is getting the attention: AI-assisted bug hunting will harden well-maintained software fast enough that within two years agencies genuinely go dark, and the demand for legislated backdoors returns in earnest. The more useful line is the one underneath it. The peace of the last ten years was never a policy outcome.
But I fear that this was less of a principled pause, and more of a market that was just pricing supply.A Few Thoughts on Cryptographic Engineering
Read that against what actually happened. Comey announced Going Dark in 2014 and asked for a national conversation. The conversation never came. GrayKey did, for locked phones. Pegasus did, for remote access. Agencies stopped pressing for exceptional access because they could buy the outcome instead, vendors kept closing holes, and offense kept its edge on the margin. Everyone in security has since treated that stalemate as a settled question — the technical case against backdoors won, and the political appetite faded. It didn't fade. It got satisfied.
It didn't fade. It got satisfied.
What changed is four months old. Anthropic shipped Mythos in April, the government briefly restricted its export, and then OpenAI, Z.ai and Moonshot demonstrated that vulnerability finding is not a capability any one lab holds. Defenders are now running the same class of model inside CI, against decades of backlog, before code ever reaches a human reviewer. Note the timing on the other side of the ledger: the White House spent Thursday standing up a program to have vetted private firms hack criminal infrastructure abroad — a capability whose core input Green expects to be largely gone by 2028.
The two-year clock is probably wrong, and the strongest objection says why: the bugs these models surface cluster into specific classes that reflect systemic architecture problems, and fixing those means overhauls that codebases assembled over thirty years will not absorb on that schedule. Granted. The ceiling arrives unevenly, and most software stays soft for a long time. But the backdoor question has never tracked the average. It tracks a short list — iOS and Android storage, iMessage, Signal, WhatsApp — all well-funded, actively hardened, and first in line to hit the ceiling. The truce breaks at the top of the distribution, which is the only place it was ever load-bearing.
So the argument comes back, and it comes back against an agency that actually cannot get in, rather than one that would prefer not to pay. That is the version nobody has had to answer. In 2016 the security community didn't win it. It just never had to have it.