Kai Tanner's filing on CVE-2026-85706 (a GitLab path traversal bug scoring 10.0 on CVSS, the ceiling of the scale) and Mei Chen's filing on Anthropic's threat intelligence disclosure are, structurally, the same story told from two ends. One is the tool a state-linked group used to automate reconnaissance and exploitation planning against Taiwan's defense networks at a volume no human team matches. The other is the kind of exposed repository that reconnaissance is built to find, and HKCERT's bulletin the same week bundles the GitLab flaw with separate bugs enabling remote code execution and cross-site scripting on the same platform. A bank or broker running a self-hosted GitLab instance is not choosing between a patch backlog and a nation-state AI operator (it is facing both, and the second one scans faster than the first one patches).

Look, the instrument that actually catches this is the SFC's 26EC32 circular on AI-enabled cyberattacks, which folds AI-specific threats into the existing cybersecurity and incident response frameworks for licensed corporations, licensed virtual asset service providers and associated entities. Anthropic's own account, that it banned the accounts involved and briefed government partners, is the kind of after-the-fact notice 26EC32 exists to make routine rather than exceptional: a firm's incident response plan has to assume the reconnaissance phase is now automated and cheap, not that a human operator will take weeks to find the exposed CI/CD credentials. For an SFC-licensed firm running GitLab internally, the obligation is not a new filing, it is proving the existing framework already accounts for AI-speed scanning against exactly the kind of unauthenticated file-read bug CISA flagged this week.

There is no consultation close or hearing date attached to either story this week; the next dated event is whenever HKCERT or the SFC issues a follow-up bulletin naming exploitation in the wild. Until then, the read stands on the arithmetic already public: a perfect-score vulnerability and a disclosed AI reconnaissance campaign surfaced in the same seven days, and 26EC32 already covers the scenario neither disclosure named directly.