A separate national security moment
The Pentagon branded Anthropic a security risk for refusing surveillance and lethal-weapons use of Claude — then carved its most-gated model out of the ban so the NSA could keep it.
Anthropic gates Claude Mythos as too dangerous to ship: the model that, pointed at the world's software, surfaced more than ten thousand unpatched security holes in weeks went out only to a defenders' consortium, withheld from the public precisely because nobody can stop an attacker from aiming it the other way. The Financial Times now reports it is being aimed exactly the other way — by the U.S. government, with Anthropic's own staff in the room.
The FT puts roughly half a dozen Anthropic engineers inside the NSA, adapting Mythos to infiltrate networks in China and Iran — though no source confirms whether they touch live operations or only customize the model.
The on-record part is the contradiction. The same Department of War that designated Anthropic a 'supply-chain risk' — a label built for adversary-linked firms like Huawei and ZTE, never before turned on an American company, the trigger being Anthropic's refusal to allow mass surveillance or fully autonomous weapons — quietly exempted Mythos from its own blacklist so the NSA could keep using it. The department's CTO, Emil Michael, confirmed the split publicly, calling Mythos 'a separate national security moment.' Anthropic is fighting the designation in two federal courts while that exemption stands.
The unconfirmed part is the escalation. Citing two people familiar with the arrangement, the FT reports roughly half a dozen Anthropic engineers embedded inside the NSA, adapting Mythos for offensive operations — infiltrating networks in countries including China and Iran. No source confirms whether those engineers touch live operations or only customize the model; the reasoning one source gave for pointing it at adversaries was simply the assumption that they already have something like it. Treat the engineer count and the targets as one outlet's anonymous sourcing, not settled fact.
What is settled is the precedent. A lab that built its brand on holding the safety line — the holdout that took a national-security blacklist rather than drop its red lines — has, on the record, let its most-restricted model be carved out for state use, and per the FT has its people helping run it. This is the template by which every frontier lab's 'red lines' meet classified demand: not a public reversal, but a quiet exemption the maker can keep fighting in court while it holds.
The lenses
The facts
How this connects
Tap a node to open it