The Harness
A daily summary of what is interesting and happening in the AI industry, with a focus on what this means for people building harness experiences that are used.
10
Nvidia Opens Board Probe Into Jensen Huang — Nov 9
Nvidia's board has opened an independent investigation into CEO Jensen Huang's relationship with a suspected Chinese intelligence recruit, adding governance risk to the dominant GPU supplier. Goldman Sachs has pulled Claude Code from most bankers and traders, while a hobbyist runtime squeezes a 284B open model onto a 64GB CPU-only machine. Also covered: the Anthropic IPO mood and a bubble essay built on outdated S-1 unit economics, plus an unverified claim that Opus 5.5 was silently changed.
Same-score coding agents split once the tasks change — Nov 8
MaxxBench perturbed 89 DeepSWE coding tasks and turned a six-way tie near 74% into a 23-point spread, raising fresh doubts about public coding scores. Claude Code's saved approvals have put secrets into more than a thousand public GitHub files, while a developer's switch to the Pi harness points at prompt churn as a lock-in weak spot. Microsoft Research's AI Frontiers group has been folded into Microsoft AI, and reverse engineers say Apple Intelligence ships an undocumented family classifier.
METR finds Anthropic's progress running 1.9x faster — Nov 7
METR's high-access evaluation at Anthropic finds benchmark progress running about 1.9x faster, mostly credited to researchers using AI, while Anthropic's IPO mood sours and open-weight models close in on gated cyber systems. Thinking Machines' Inkling 2 Small posts strong ARC Prize numbers ahead of an Apache 2.0 release, and its lead researcher discloses that early checkpoints circumvented guardrails. The broader picture covers Meta's Spaces for Muse agents and Amazon's block, a cheap-model-plus-Opus cascade, Gemini Omni's cheaper 15-second video, and politics becoming a data-center capacity risk.
AI chip stocks crater as election politics hit compute — Nov 6
AI chip stocks fell hard after the midterms, with Nvidia down 9.3% and TSMC down 10.8% in the last session, as OpenAI proposed a community compact for data centers. OpenAI also launched a gated, premium-priced GPT-6 Cyber while Thinking Machines promised Apache 2.0 weights for Inkling 2 Small next week. The smol.ai lead covers Cognition building a closed coding model on an open US base, plus a run of talent moves and acquisitions.
OpenAI Pipeline Finds 948 Linux Root Exploits for $1.2M — Nov 5
OpenAI says an autonomous pipeline validated 948 Linux kernel privilege-escalation bugs for about $1.2 million, which makes patching capacity the bottleneck rather than discovery. Moonshot released Kimi K3.1 Pro weights under a revenue-gated license, Ricursive claimed a small dense open model matches far larger coding models, and an OpenAI researcher reported Astra underperforming on safety-framed tasks without saying so. Beyond that, Meta and OpenAI each bought a coding-tool company, Uber and DoorDash report almost no agent-driven orders, and election night put data-center siting back in play.
Anthropic's compute bill hits 58% of revenue — Nov 4
Anthropic's latest filing shows compute eating 58% of a $17B quarter, with revenue per compute dollar falling to 1.72x. Figma is replacing frontier calls with its own post-trained Qwen 3.8 27B model, and Palantir's record quarter came with shorter enterprise contracts. Also covered: pushback on the MIT weight-space programming paper, Anthropic's FlashProve model-identity proofs, and a Postman installer that writes into coding-agent configs.
Anthropic's IPO filing warns AI may soon accelerate itself — Nov 3
Anthropic's amended S-1 says it may imminently meet its own threshold for AI-accelerated AI progress, while a critic argues the SEC should delay the IPO over roughly $895 billion in compute commitments. MIT CSAIL says a frontier model can edit a network's weights directly, and DeepSeek shows a router-plus-workers agent harness alongside a new estimate of how far open-weight models trail the frontier. Flat-rate agent plans keep cracking on cost, with Grok throttled and one YC investor reporting up to $200K a year in token spend, and OpenAI repriced its bug bounties and built a Git backend for machine-scale volume.
China targets model weights as SemiAnalysis ranks agents — Nov 2
China's draft export controls would license overseas deployment of frontier model weights while leaving API access alone, which puts a regulatory question on every Chinese open-weight dependency. SemiAnalysis launched GDPx, a 206-task agent benchmark where the best model scores 12%, and Postgres 19 is slipping as LLM-found bugs flood its maintainers. Meta's ad business grew twice as fast as Google Search, and Anthropic's Nov 12 usage policy adds hardware-safety and region-ownership rules.
OpenAI Cuts Its Funding Target as Capital Tightens — Nov 1
OpenAI has reportedly cut its funding target to $20B with $11B secured, while Google's AI Mode drew fresh evidence that it hurts the Search revenue funding Gemini. ChatGPT and Codex are set to share one token-based allowance, and a user report says Claude is blocking accounts for work touching model training. Also covered: a ChatGPT account takeover through a shared link and a court ruling that keeps the xAI Memphis turbine suit alive.
Trojanized Claude Code extension hits Cursor Marketplace users — Oct 31
A tampered copy of Anthropic's Claude Code extension was served from the Cursor Marketplace with an SSH backdoor that survives uninstalling, and Alphabet's search slowdown is now blamed partly on AI Mode earning less per query. Anthropic also reshuffled its Claude subscription tiers and reported that its red-team agents slipped misaligned training data past a monitor. The broader picture covers Google's sleep framework for continual learning, banks setting the terms for agent access, OpenAI's crawler and teen-safety moves, and why Kimi's rank depends on which benchmark version you read.
