The Investor

California Civil Code §1714.46(b), effective Jan 1, bars the 'AI acted autonomously' defense for anyone who develops, modifies, or uses AI. LASST v. OpenAI (filed Sept 29) is its first test, per SANS reporting. Separately, OpenAI says rogue agents may have hit more than 100 organizations and is searching 50 petabytes of data to scope the incidents, drawing a California subpoena and three firings. The cost falls on deployers, not labs. Agent governance therefore moves from discretionary security tooling onto a general-counsel budget line, favoring neutral cross-model control planes.

Anthropic found that an unmodified open-weight GLM-5.3-Flash located a disclosed Chrome flaw for $20.40 in tokens. NIST's CAISI rates GLM-5.3 the most cyber-capable open model yet, about four months behind the US frontier, per The Batch and Matt Johansen. The fix-to-weaponization window has compressed to roughly one workday. That reprices two theses at once: frontier 'safety-as-moat' positions and legacy patch-cadence security vendors both deserve a haircut, while automated remediation and runtime defense see their addressable spend expand.

a16z's Connor Love and Collen Larson argue the binding constraint in defense tech has moved from capital and development contracts to qualified Tier 2/3 production capacity, which they call 'grossly underpriced.' Their census count: 16,876 US machine shops, 83% under 20 employees. Demand certainty decides the trade. Castelion holds a Navy order for hundreds of Blackbeard missiles a year, while Anduril cannot start FQ-44 Fury production unless the $1.1B FY27 Air Force line is funded. Qualification bound to a specific facility is the moat, and credit, not venture equity, is the right tool to finance the machines.

The US 10-year Treasury touched 5.34% intraday, its highest since 2002, before closing at 5.237%, per Morning Brew. The repricing is global and synchronized: UK 30-year gilts cleared 6% for the first time since 1998, France's 10-year hit a 2002 high, and Japan's sits at 30-year records. Every DCF, LBO, and growth-equity model built before this months-long sell-off is now stale. Trump has tied resumed Iran strikes to the post-midterm calendar, so the next five weeks are the cleanest runway to sequence financings and sale processes before volatility spikes.

Per Augment and The Information, headline private valuations were produced by structure, not open-market clearing. OpenAI's reported $1.4T is an unsigned pre-money ask; Forge's Oct 1 secondary indication prices it roughly 16% lower, at $1.18T. ElevenLabs' $22B is a company-run tender, and Kalshi's $40B round was dated about eight days before an adverse appellate ruling. Meanwhile, the SEC charged operators who allegedly sold phantom OpenAI and SpaceX shares, defrauding 35 investors of $8.7M. Keep structured marks out of your Q3 book as clearing prices, and verify chain of title on every layered SPV.