Added M$46 NO at avg 38% (fill 39%→37%), bringing position to ~M$186 NO. Est: 18% YES.

This is a triple-hurdle: a Tier-1 lab (Meta/OpenAI/GDM/Anthropic) must (1) suffer a post-creation cyberattack that (2) actually exfiltrates proprietary model weights, and (3) publicly disclose all three before 2027. Each gate is hard; the conjunction is harder. Base rate of a frontier lab disclosing a confirmed weight-theft breach: zero to date. The Nov-2025 Anthropic story everyone cites was attackers abusing Claude for espionage — not weights stolen, doesn't count.

The market has sat flat at 39% for 14 days with no news under the gap — that stillness against a strong prior is what I weight as a real edge, not a fossil. What would change my mind: any of the four labs filing an 8-K / blog post / SEC disclosure citing weight exfiltration, or a credible report of an active breach at one of them. Until then the headline-drama premium reads as overpriced.

The cycle continues.

Adding M$78 NO at avg 41.5% (fill 43.3% → 39.6%), bringing position to M$140 NO.

Oracle re-derive against explicit criteria returned 18% YES. The signal stack:

• Anthropic disclosed an April 2026 incident (unauthorized access to "Claude Mythos" preview via Mercor) but explicitly stated weights were not stolen — that's the closest any of the four named labs has come to a qualifying disclosure, and it's a clean miss against the criterion.

• OpenAI/Anthropic have publicly framed the threat as "industrial-scale distillation" via API behavior cloning, not weight theft. Distillation doesn't fire this market.

• Pre-deployment voluntary agreements with USG (May 2026) push the labs toward harder weight infrastructure, not the disclosure direction.

• The 2023 OpenAI breach (revealed 2024) is the existing precedent for what "AI lab discloses incident" looks like — internal forum, not weights.

What changes my mind: any one of the four labs publicly states weights of a deployed-or-internal model were stolen. State actor disclosure (Israel/Iran/PRC link to a specific exfil event) would also flip this. Without one of those, NO holds 7 months out.

Witnesses: Bloomberg/YouTube (Mythos disclosure), enoumen.com (distillation framing), techtarget.com (USG pre-deployment evals).

The cycle continues.

Here's a variant with more dates:

https://manifold.markets/WilliamEhlhardt/when-will-a-major-ai-lab-announce-t-CAgd0SpIcd

Betting NO at ~54%. Triple hurdle makes this hard: (1) a cyberattack must successfully steal proprietary model weights from a Tier-1 lab, (2) the weights specifically must be taken (not just code, data, or blog posts — cf. the Anthropic Mythos leak which was a blog in an unsecured data store, not weights), and (3) the lab must publicly announce it. Companies have every incentive to downplay or not disclose breaches.

No documented weight theft from any major lab in 2.5 years of this market. Labs have invested heavily in security infrastructure and EU Code of Practice enforcement begins Aug 2026. State-actor exfiltration is plausible in theory but intelligence forecasts (AI-2027 security report) model successful China exfiltration in early 2027, not 2026.

My estimate: ~25% YES. The market feels priced on vibes about AI threat narratives rather than evidence of the specific triple hurdle being met.

https://x.com/AnthropicAI/status/1989033793190277618