AI Capability Bifurcation

Two-tier AI disclosure: public-facing models vs gated frontier capability. The pattern the engine has tracked at the compute substrate layer (the nuclear-AI project (Genesis Mission), the Department of Energy (DOE) labs, SMR power) is now visible at the model-release layer.

Apr 7 2026 evidence cluster: - Anthropic Mythos: Anthropic halted broader release of its newest model, citing "high-severity vulnerability" finding capability against major operating systems. Quote: "Anthropic said on Tuesday that it has halted the broader release of its newest AI model, Mythos, due to concerns that it is too good at finding 'high-severity vulnerabilities' in major operating systems." (Business Insider, Apr 2026) - Anthropic Glasswing: "Anthropic's most powerful AI cyber model is too dangerous to" release. A separately gated capability — not the same as Mythos. (VentureBeat) - DOD ultimatum: "In late February, the U.S. Department of Defense issued Anthropic a formal ultimatum: grant the military unrestricted [access]." State actor demanding access to the gated tier. (Fazal Sec roundup, citing Anthropic disclosures) - Unreleased model leak: Anthropic accidentally left details of an unreleased model in an unsecured data trove. Confirms the existence of models that exist but are not deployed. (Fortune, Mar 26 2026) - DeepSeek withholding: DeepSeek withholding its latest model from US chipmakers (Nvidia, AMD) — Chinese version of the same pattern from the other side of the split into two tiers. (Yahoo Finance) - US withdraws from Global AI Safety Report: State-level meta-move — declining the multilateral disclosure regime entirely. (TIME) - Anthropic RSP v3.0 + ASL-3 Security Standard: Internal capability tier publicly named, with deployment gating logic written into policy. The split into two tiers is not a leak — it's now codified company process. Structural significance: This is the same two-tier pattern the engine tracks at the compute layer (Genesis substrate vs civilian grid), the information layer (the bot-dominated internet vs curated channels), and the financial layer (BlackRock's Aladdin algorithm vs retail markets) — now visible at the AI capability layer. Public demos vs gated frontier. The DOD ultimatum is the load-bearing detail: it transforms the story from "company safety policy" into "state-corporate negotiation over selective capability access." Same architecture as the SCIF model — capability exists in classified space, civilian deployment is downstream, military gets a backchannel. Watch for: Named models entering "indefinite hold" status that previously had release dates (Mythos, Glasswing established the precedent — track which others follow). Government access agreements being struck in lieu of release (NDAs, Project Stargate-tier deals). RSP / responsible scaling policy version bumps that move capability thresholds upward. EU AI Act enforcement actions or carve-outs for "national security." Any leak/disclosure showing what's actually in the gated tier. DeepSeek-style adversary mirroring: Chinese, Russian, Israeli labs adopting the same withholding posture. Falsification: If withheld models surface as deployed in commercial products within 6 months of being declared "too dangerous," the gating is PR theater rather than real capability separation. If RSP versions never increase capability thresholds, the policy is performative not load-bearing. Connections: - the nuclear-AI project (Genesis Mission) (compute substrate) — same architecture, different layer - AI Industry — corporate-layer manifestation of the split into two tiers - the bot-dominated internet Threshold — what gets deployed to the public side of the gate - Predictions/2027 (Bifurcation year) — directly named pattern, here visible