AI News: Thursday, October 8, 2026

The White House says eleven companies pledged $2.4B in compute credits and AI tools to the Genesis Mission; Nvidia ($1B) and AMD ($500M) lead, and the amounts are pledges, not cash.

USA Today Co. sued OpenAI on October 8, 2026, seeking over $250 million for allegedly copying hundreds of thousands of articles to train AI models; OpenAI had not responded yet.

Anthropic OSS Scanner scans enrolled open-source repos with its strongest models for free and emails unreviewed reports with reproducers and patches, with no 90-day public disclosure clock.

Scott Aaronson cites anonymous sources saying AI labs are testing whether models can break cryptography; no break has been reported, and Ethereum researchers are urging precaution, not panic.

Goodfire launched activation-probe monitors for Baseten customers that flagged 94% of malicious hacking sessions in tests on Kimi K3, at roughly $51 per 1,500 sessions versus about $10,000 for a top-tier LLM judge.

OpenAI banned a Russian and an Iranian influence operation that used fake journalists and a fake think tank; the Russian one is its first Category 5 case on the 6-point Breakout Scale.

Zenity Labs says one prompt to a public AgentCore agent exposed its metadata-service credentials, and a broad default role then let attackers reach every agent in the account; AWS has partly fixed it.

Anthropic launched the Cyber Mission: a Critical Infrastructure Defense Program giving 11 founding partners frontier Claude models, engineers and threat research for OT security, alongside free OSS Scanner.

OpenAI confirmed it fired three safety researchers for mishandling sensitive information reportedly shared with an outside AI-safety group; the recipient is unnamed and key details are unpublished.

Guardian Australia reportedly found OpenAI used AI to help draft its September 10 breach notice, though Jason Kwon told an inquiry he did not believe so; a source says humans reviewed it before sending.

The 81% figure measures task exposure to AI plus robots, not jobs replaced; Anthropic's robot study finds robots cost-competitive for only about 0.3% of tasks, so hiring is not a contradiction.

SemiAnalysis counts over 24 GW operating in China plus about 50 GW in the pipeline versus about 56 GW in the US; the buildout is gated by chips rather than power, so GW overstates effective compute.

Microsoft says Windows 11 will route AI work between local hardware and the cloud, with Execution Containers for agents now generally available and local Copilot features arriving over the coming months.

Tab is a text-message personal assistant with its own phone number, computer and wallet that emerged from stealth at a $300M valuation; its privacy promises are claims, and the amount raised and pricing were not disclosed.

ICANN published 1,615 applications for new top-level domains on October 7, 2026; press reports say OpenAI and Meta are among roughly ten applicants for .agent, with .agi and .asi also contested.

A Hugging Face post shows its ML Intern agent training six custom models for about USD 103 in total, and shares the prompt rules (baseline, smoke test, budget cap) that kept each run on track.

Anthropic's 2026 usage policy, effective November 12, bans sustained and needless cruelty toward Claude and tightens rules on weapons, surveillance, deceptive campaigns and autonomous hardware.

Anthropic says Claude Science agents merged public UV surveys and predicted the unobserved third of the sky to within about 10 percent, producing the first complete UV map of the sky.

Perplexity's pplx-embed-v2-late keeps token-level vectors for text and images so parts of a query match parts of a document, in 0.6B and 9B sizes that share an embedding space; benchmark scores were not independently confirmed.

Sonnet 5.5 cache reads dropped from $0.20 to $0.10 per million tokens; the roughly 20% saving holds only when cache reads dominate tokens and output is tiny, while chat-style mixes save about 5%.

The Association for Human Mathematics, in a statement reposted by Terence Tao on October 7, 2026, urged mathematicians to stop working with OpenAI after its October 6 release of 700-plus math manuscripts.

A paper says on-policy distillation is RL where the teacher is an implicit reward: it sharpens what the student can already solve, and collapses into repetition when the teacher rewards the wrong behavior.