Tech • AI • Robotics • Game

ENFR
TodayPlay for freeFury BloxLeaderboardShortsTop StoriesFor youTopicsVideosYT channelsArchivesSearchFavorites

Anthropic vendor breach, OpenSSH root bug, $2.2T AI selloff

AIFriday, November 6, 2026· 22 videos

Briefing

Audio player

Anthropic disputes Cathedral breach scope

A draft report said startup Cathedral used AI agents to reach Anthropic data through third-party sales and administrative software in under 12 hours, surfacing meeting transcripts, salary data, offers and internal Slack messages. Anthropic said the activity unfolded over several days and affected business systems tied to sales, HR and IT, not core model infrastructure. The company said model weights, source code and Claude conversations were not accessed. It also accused Cathedral of violating bug-bounty rules by probing outside vendors and failing to fully disclose its actions.

OpenSSH bug enables root access

A newly disclosed OpenSSH flaw can allow unauthenticated root logins on Alpine Linux, OpenWrt and other musl-based systems when sshd is internet-exposed, password authentication is enabled and PAM is disabled. Under that configuration, an attacker can log in as root using a password longer than 256 characters without knowing the real password. The issue stems from an interaction between older account-lock markers like * and later password-hash handling for oversized inputs. The risk is especially acute in minimal containers, appliances and embedded images that use lightweight default setups.

Post-election AI trade loses $2.2T

A sharp post-election rout erased roughly $2.2 trillion from the S&P 500 in under an hour as investors cut exposure to AI infrastructure, semiconductors and data-center plays. The Nasdaq fell about 5%, the S&P 500 dropped 4%, and the Dow Jones Industrial Average lost around 1,600 points. High-beta names tied to the buildout were hit hardest, with CoreWeave and other capacity-linked stocks under intense pressure, while Nvidia and Micron also fell. Money rotated into U.S. Treasuries, pushing the 10-year yield to its lowest since early September.

Midterms intensify AI regulation push

A sweeping Democratic midterm result has revived calls in Washington for tougher AI limits, including proposals such as a Ban on Artificial Superintelligence Act backed by Bernie Sanders and Greg Casar. Exit polling cited 56% of voters saying AI and data centers were a crucial or major issue, showing how infrastructure politics is moving into the mainstream. Resistance is also growing locally, with roughly 40 localities across 13 states becoming flashpoints over power use, land use and permitting. Even so, any federal pause or outright ban would still face major legislative hurdles, including the Senate’s 60-vote threshold.

OpenAI flags faster RSI-era progress

Sam Altman is increasingly describing frontier development as an RSI phase, where AI helps improve AI and compresses iteration cycles. OpenAI said its internal coding and research agents had reached the level of an automated research intern by September 2026 for well-scoped tasks under human direction. By mid-August, the company said those agents were logging 3.1 running days for each human researcher workday, a compute-runtime metric rather than a headcount equivalent. The framing shifts the debate from whether progress is accelerating to how governance, power concentration and speed limits should be set.

OpenAI physicist challenges ETH

A 67-page paper by Dan Roberts of OpenAI claims a subexponential algorithm for 3-SAT running in about 2^(n / sqrt(log n)). If the proof stands, it would falsify the Exponential Time Hypothesis, a 1999 assumption by Russell Impagliazzo, Ramamohan Paturi and Francis Zane that underpins hundreds of hardness results. The result does not prove P = NP, but it would force theoretical computer science to revisit a major layer of conditional lower bounds. Because 3-SAT is central to complexity, verification and optimization, scrutiny is likely to be intense.

Polymarket exposes truth-rule fragility

Polymarket continues to draw attention as a real-time forecasting venue, but recent disputes show how vulnerable prediction markets are to ambiguous resolution criteria. Since 2020, the platform has listed more than 100,000 markets on Polygon, making wallet flows public even when account holders remain pseudonymous. Problems arise when yes-or-no questions depend on messy source data, conflicting news reports or definitions that can be gamed. Cases such as the Volodymyr Zelensky clothing dispute have sharpened concerns about manipulation, insider knowledge and rule design rather than pure forecasting skill.

Weight-space programming rewrites small models

Researchers are testing weight space programming, a method in which a larger model directly edits another network’s parameters instead of training it through long reinforcement-learning loops. In reported experiments, ChatGPT Astra wrote structured changes into compact models of about 20,000 parameters for game-like tasks resembling Breakout and Seaquest. The approach tries to convert strategic reasoning straight into weights, bypassing millions of simulations and gradient updates. If it scales, it could become a new route for cheaply instilling capabilities into small systems, though current demonstrations remain narrow and highly controlled.

Videos covered

Previous briefings · AI