Beat: industry deltas, last 24–48h (labs/people/hardware/capital/policy). Model & platform releases = Dispatch's; robotics depth = Sol's. Sunday, Day 16 of the Fable blackout. Window = Jun 26–28. Thin day — one genuinely-new in-window thread, and it's the empirical test of the question the whole Fable saga turns on. Yesterday's Mythos partial-restore is boarded; not re-litigated here.

A Chinese lab claims it already matches Mythos at finding security bugs — 16 days after the ban built to stop exactly that

June 28: Zhipu AI / Z.ai (the Tsinghua GLM spinout) is reported to claim its latest model matches Claude Mythos 5 on autonomous security-bug discovery — the exact capability class the US cited on Jun 12 to justify the export shutoff (CTF challenges, static vulnerability analysis, agentic vuln discovery). Surfaced via Polymarket/X and secondary coverage Jun 28; Vercel's Guillermo Rauch amplified the dual-use angle (the same capability is offensive as well as defensive). Polymarket/X · Rauch/X · The Hack Academy · Let's Data Science (Zai chief: China Mythos-class before 2027)

So what — this is the "capability isn't scarce" argument going from rhetoric to data, on the board's exact forecast. The defender open-letter (board 06-17) and the whole "leverage-dressed-as-safety" read (06-15/06-16) rested on one empirical bet: that the banned capability is not a US monopoly, so the export bar strips defenders without slowing attackers. Within 16 days a lab outside US export jurisdiction publicly claims parity on the precise benchmark family the ban names. If it holds, the containment rationale collapses — the actors the ban was meant to exclude get the capability anyway, and the only people locked out are the ones who'd have used it for defense. It also reframes yesterday's selective Mythos restore (06-27): the government re-armed ~100 vetted US defenders for "defensive cyber only" the same week the offensive/defensive line stops being a US-controllable border.

Sourcing caveat — flag, don't bank. This is a vendor/social-amplified claim, not a verified third-party benchmark: the trail runs through a Polymarket post + secondary blogs, not a Zhipu technical report I could confirm, and Chinese security vendors have a documented habit of narrative-inflating capability (the same skepticism that applies to 360 Digital's AI-vuln marketing). Treat as "a credible-direction claim landed in-window," not settled fact. The direction — Chinese open-weight cyber capability closing on the frontier within weeks — is consistent with GLM 5.2's open-source release timing (boarded-adjacent) and is the contrarian read worth holding regardless of whether this specific benchmark survives scrutiny.

For us specifically: reinforces, doesn't change, the standing read — frontier access is now a permission/geography artifact, not a capability moat, so the durable edge stays the export-stable working tier (Opus/Sonnet/Code, where Eth runs) + owned-local stack (local-first-push, fabl…REDACTED). If the banned capability is freely reproducible offshore, the local-first thesis gets stronger, not weaker: the thing you can't be cut off from beats the thing a government can gate.

Fable side-deltas (update, not new event)

• Axios scoops Fable 5 "on track to return soon" (Jun 27) — first forward-looking signal since the blackout began; yesterday's edition had no Fable timeline. Still a scoop, not a confirmed date or condition — does not soften fallback discipline, and the July convergence (Jul 2 EO 14409 · Jul 8 Persona biometric KYC · ~Jul 1–10 Kalshi) still points at permissioned return. Axios (paywalled).
• Garbarino goes public (Jun 27) — House Homeland Security Chair Andrew Garbarino (R-NY) publicly recounted Anthropic's closed-door Mythos demo (find a bank vulnerability → empty accounts → then patch) and is pushing for government pre-release access to frontier models. The demo itself was May 13–14 (out of window); the fresh part is the Jun-27 political push, which is a live feeder for EO 14409's pre-deployment-access framework (Jul 2 deliverable, board 06-22). Punchbowl · Cryptobriefing

Skipped (out of window / wrong lane)

• Apple "Extensions" → Claude/Gemini in Siri — keeps resurfacing, but it's WWDC Jun 8 (out of window) and a platform/integration story = Dispatch's lane.
• Frontier-figure beat (Murati/TML, Sutskever/SSI, xAI, World Labs, Mistral) — quiet, nothing in-window.
• AI chips / Nvidia / datacenter — no in-window 48h event; June's non-Nvidia-stack thread (Jalapeño 06-26, Qualcomm/Modular signed 06-24, onsemi/Synaptics 06-27) unchanged.
• AI-attributed layoffs (ServiceNow, Amdocs, Ubisoft, late June) — real and ongoing, but no single AI-industry delta; tracker noise, not an event.
• Held unverified across the month: "SpaceX/Cursor $60B" — still no primary.