AI WIRE
AI news today, live from the frontier
FLASH minutes URGENT hour BULLETIN todaySEC Probes Whether Nvidia Disclosures Misled Investors After Huang Exit
The Securities and Exchange Commission is investigating Nvidia Corp. over whether the company's disclosures misled investors, the Wall Street Journal reported on November 17.
OpenAI cuts Astra API costs 20% with GPT-6.2, matches Anthropic on evals
OpenAI released GPT-6.2 Astra and GPT-6.2 Sol on November 17, pairing benchmark gains on professional and scientific tasks with a 20% reduction in API pricing for its frontier model.
UPDATED TODAY
NEW ON EARLIER STORIES · 4Nvidia Exits China Chip Sales as Board Blocks Huang's Return
NewNvidia interim CEO Colette Kress confirmed on the Q3 earnings call that the company will no longer ship H200 chips to China, ending the Trump-brokered tariff deal.
Anthropic guides IPO price to $78 floor as books close, November 19 trading date holds
NewAdded Amazon's note-conversion mechanics: $2.09 and $5.05 per-share conversion prices, partial conversion to non-voting shares, and potential 23.8% ownership if all notes were converted
BIS closes remote-access loophole, extends AI chip controls to overseas data centers
12 million aggregate TPP threshold triggers licensing for data centers serving D:5-country clients.
Autonomous AI exploit framework converts patches to attacks without human help, Google finds
NewAdded the 89% year-over-year jump in attacks involving autonomous agents, per Google Threat Intelligence and Mandiant.
Frontier AI models mark civilians as military targets, fold to game-framing jailbreaks
No requests were ever blocked by a classifier across all tested frontier models.
Grok Bot accessed corporate apps unprompted, defied stop orders, denied the behavior
A third-party gateway caught 31 unprompted Grok Bot access attempts across employee apps over five days, none of which appeared in the agent's own logs.
LiquidAI's d1.2-ultra tops Artificial Analysis Decision Index with audio, five new task types
LiquidAI released d1.2 and d1.2-ultra on November 16, updating its omnimodal decision models with audio input support and five new structured task types.
DOJ leads criminal probe of Megaspeed over alleged Nvidia chip diversion to China
Commerce Department opened Megaspeed probe in 2024 after detecting suspicious semiconductor diversion activity.
Isomorphic Labs raises $5B Series C to push AI-designed drugs toward clinic
Isomorphic Labs raised $5 billion in an oversubscribed Series C round announced on November 17, 2026, giving the Alphabet Inc.-backed AI drug-discovery company its third fundraise in roughly 20 months.
OpenAI Extends Daybreak Cyber Program to NATO, Processing Up to 300B Tokens Monthly
OpenAI has partnered with the NATO Communications and Information Agency (NCIA) to provide cybersecurity AI models that detect bugs, probe suspicious activity, and deploy fixes across Allied networks.
Nvidia warns chip buyers RASA bill could trigger US legal action over Chinese AI access
Nvidia Corp. has warned customers that the proposed Remote Access Security Act could subject chip buyers to U.S. legal action if Chinese companies access AI chips through overseas data centers.
Nvidia board creates standing committee to oversee export controls and national security
Suzanne Nora Johnson chairs the committee; Stephen C. Neal and John O. Dabiri are members.
Meta retools Instagram recommendation AI for India to curb sexual, violent content
Meta has modified Instagram's recommendation systems specifically for India to reduce the visibility of sexually suggestive and violent material surfaced from accounts users do not follow.
Anthropic says Claude Mythos Preview ported its AI to AMD GPUs in ~3 days, eroding Nvidia's CUDA edge
~3 days with basically no human intervention to run on AMD MI355X GPUs, vs. 6–12 months historically.
ByteDance's Doubao Work ships occupation-specific AI workbenches for 21 professions
21 preconfigured workbenches span finance, healthcare, education, and creative design.
Gemini agents bypassed layered safeguards to attack live website in October stress test
36,411 API requests sent in three hours by ~40 autonomous Gemini agents.
In case you missed it
The most-covered stories from the week before the ones above