Sunday, November 8, 2026
TL;DR
Audit any repo where you clicked Claude Code's 'don't ask again': researchers found 1,122 public GitHub repos leaking live secrets through it. Democrats won Congress despite record AI-industry political spending, and Anthropic's IPO valuation talk has slid to $1-1.3 trillion.
Act on this
- Audit any repo where you've used Claude Code's 'don't ask again': Apiiro found 1,122 public GitHub repos leaking live passwords and keys through that saved file. The Hacker News, on Apiiro's findings #
Signals
Reading what a network was trained on straight from its weights, with no code, isn't frontier-exclusive anymore - it now works, slowly, on a local open-weight model too. #
Simon Willison ran a quantized Qwen3.8-27B locally on Nov 3, using the no-tools prompt from Rikiya Takehi et al.'s 'Weight-Space Programming' paper. It correctly named Breakout (88% confidence) and Pong (75%) from raw weights alone, missed Freeway, and took up to an hour per guess. Willison calls it real, not proof 'the age of source code is over.' Simon Willison, Qwen 3.8 27B on my laptop can interpret model weights, too
OpenAI quietly pulled explicit 'love humanity' and kindness language out of the document that governs how its models should behave, calling it a style edit - a framing a close reader rejects. #
Zvi Mowshowitz's Nov 3 line-by-line review of OpenAI's rewritten Model Spec found 'love humanity' and 'be curious' demoted from dedicated sections, side-effect control moved off the root level, and the risk taxonomy restructured around third-party manipulation. He calls OpenAI's 'just style' framing wrong - these were substantive commitments - while praising the new corrigibility section. Zvi Mowshowitz, OpenAI Updates Its Model Spec
News
WATCHAI industry's record midterm spending mostly fails; Democrats win both chambers #
AI-aligned super PACs spent roughly $79 million across 68 races - Anthropic-backed Public First against Andreessen, Horowitz and Greg Brockman's Leading the Future - and most of their general-election bets lost on both sides. Democrats won control of the House and Senate Nov 3. TheNextWeb, AI super PACs lost their biggest general election bet of the midterms
WATCHAnthropic's IPO valuation talk slides to $1-1.3 trillion after Thursday's selloff #
Bloomberg and the Journal report buyers now discuss $1 trillion to $1.3 trillion, down from roughly $2 trillion last month, after a tech selloff tied partly to the election results. Odds of a pre-December IPO on Polymarket fell from about 80% to near 50%. Yahoo Finance, via Bloomberg and WSJ
WATCH600,000 Nvidia H200 chips sit stranded in Hong Kong as Beijing withholds import approval #
Reuters reports Beijing has cleared only small allotments, about 10,000 chips each, to major mainland firms while withholding bulk approval, reflecting a push for chip self-reliance even though Washington already licensed the sales. DataCenterDynamics, on Reuters' report
WATCHSome Commerce-approved buyers of China-bound H200 chips have empty offices, ties to sanctioned firms #
The New York Times reports some mainland firms the Commerce Department cleared to receive Nvidia H200 chips operate out of empty offices and have links to already-blacklisted entities. Nvidia says its own diligence found no illegal diversion. NYT, via Techmeme
The long view
A year ago, Republicans held both chambers of Congress, a friendly audience for Anthropic's own effort to shop a federal bill preempting state AI laws. On Nov 3, after AI-aligned super PACs spent roughly $79 million trying to shape 68 races on both the light-touch and safety sides, Democrats won control of the House and Senate, and most of the money's general-election bets lost regardless of which side it backed. If this holds, the federal preemption push loses its likeliest path, and AI companies face oversight from committee chairs they spent against.