Wired: Cyber-Capable AI Will Spread Despite Anthropic Curbs

The US government's emergency move to pull Anthropic's most capable models off the open market lasted about as long as it took a Chinese lab to ship a free alternative, and that compressed timeline is the heart of the Wired argument. Locking down one cyber-capable model does not lock down the capability.

The factual sequence is tight. On June 12, 2026, the US Commerce Department directed Anthropic to suspend access to Fable 5 and Mythos 5 for any foreign national, including the company's own foreign national employees, after citing a method of bypassing the model's safeguards. Anthropic complied. The next day, Zhipu AI released GLM-5.2 under a permissive open-weight license, meaning any researcher or developer could download and run it on standard consumer-grade hardware. Semgrep then reported that GLM-5.2 scored 39% F1 on IDOR vulnerability detection, beating Claude Code at 32%, at roughly $0.17 per vulnerability found. Graphistry's separate evaluation reached a similar conclusion on cybersecurity investigation benchmarks.

Why that matters for anyone doing security work, on offense or defense, is that the policy lever the US reached for — export controls on a closed API — has no neat analog for a mathematical file released globally under an open license. Closed weights can be revoked, jailbreaks patched, employees screened. Open weights, once downloaded, can be fine-tuned offline, stripped of guardrails, and run with no visibility back to any provider or defender.

The honest caveat is that benchmark parity on narrow tests like IDOR is not the same as full operational parity in a real intrusion, and the US position is already moving. Commerce Secretary Howard Lutnick has since written that appropriate safeguards are in place to permit certain trusted partners to access Mythos 5, though the same letter did not include permission for Anthropic to release Fable.

The forward read is less about which single lab ships the next 'dangerous' model and more about the cadence. If a roughly equivalent open-weight release keeps landing every few weeks, the defender's job shifts from gating access to assuming the capability is already in the room.