AI News
Latest AI engineering news, updated daily.
Ai Engineering
John Langford says Microsoft's AI Frontiers is over, and blames a scaling-only Microsoft AI
In a November 6 post, the co-founder of the roughly 50-person lab behind AutoGen and Fara says it was reorganized into Microsoft AI in September and that he has left. Microsoft has not given its own account.
Microsoft · Microsoft Research · Ai Labs
Ai Agents
ThisBash: a browser extension can drive Claude Desktop through Dispatch, and the flaw is still open
Forever Security says any extension that controls claude.ai can make Claude Desktop run code, even on another machine on the same account. Anthropic closed the May report as a duplicate, and its planned October retirement of Dispatch has not happened.
Claude Desktop · Agent Security · Browser Extensions
Ai Engineering
Nvidia's board opens an independent probe after reports Jensen Huang was a Chinese intelligence target
On November 8, Nvidia's independent directors launched an investigation and Huang admitted an affair. The harder questions concern Megaspeed, chip diversion and what the Trump administration knew, with earnings on November 17 and a House briefing the next day.
Nvidia · Export Controls · Ai Chips
Ai Engineering
Editing 1,023 weights raised Qwen3.5-0.8B's MATH-500 score by 29 points. Most of the paper is about why that rarely works
A new paper from Shanghai AI Lab, Tongji and Fudan tests agent-written weight edits across 16 trajectories. The headline gain is real on held-out data, but six runs kept the original weights and the larger model barely moved.
Weight Space Programming · Interpretability · Model Editing
Ai Engineering
The New York Times and three Tribune-group papers sue Brave, targeting the search API that feeds AI companies
Two November 2 complaints in Manhattan federal court accuse Brave of selling news through its Search API and Ask Brave, bypassing paywalls and replacing publisher ads. The tests cited show near-complete articles returned through an endpoint built for LLMs.
Copyright · Brave · Rag
Ai Engineering
DOJ's Super Intelligence Task Force absorbs the unit created to fight state AI laws
The Justice Department announced DOJ-SITF on November 5. The two-page memo behind it merges the January AI task force, tells prosecutors to give notice on request, and never defines Super Intelligence.
Ai Regulation · Us Policy · Doj
Ai Coding
Apiiro finds real secrets in 1,122 public Claude Code settings files, and 215 of them evaded two scanners
Choosing "Yes, and don't ask again" can save a command's password verbatim in .claude/settings.local.json. Apiiro counted 2,539 distinct secrets across 1,994 public repositories.
Claude Code · Secrets Management · Supply Chain Security
Ai Agents
DeepMind's AI agents cut the theoretical cost of factoring many 1024-bit RSA keys, ending a 33-year stall
A November 5 working paper gives each further 1024-bit integer about 1/58 the cost of a number field sieve run after one run of precomputation. No fan factory has been built, and RSA-2048 is not threatened.
Cryptography · Rsa · Google Deepmind
Ai Engineering
Thinking Machines' Inkling 2 Small jumps 18 points on Artificial Analysis, with Apache 2.0 weights still pending
The 594B-parameter, 17B-active model launched November 5 on Tinker with closed weights. It scores 31 on the Artificial Analysis index, but uses about four times the output tokens of its predecessor.
Thinking Machines · Open Weights · Mixture Of Experts
Ai Agents
Anthropic maps the agent memory stack: dreaming lifts wiki recall from 13% to 48%
Anthropic's new engineering post breaks agent memory into transcripts, notes, wikis and injected context, and shows with Pokémon, a 420M-token codebase and legal briefs where idle-time 'dreaming' helps and where it does not.
Agent Memory · Anthropic · Claude
Ai Engineering
Nous Research's Arete-27B cuts Qwen's state-line answers from 22% to 0.9% with SFT alone
Nous Research released Arete-27B, an Apache 2.0 fine-tune of Qwen3.8-27B that it says strips Chinese state-aligned framing while keeping capabilities. The numbers are self-reported, judged by Grok 4.7, and the method is cheap enough to matter.
Open Weights · Fine Tuning · Nous Research
Ai Engineering
PSI says a non-public Anthropic model proved the Aaronson-Ambainis conjecture, with a full Lean check
Physical Superintelligence says its research system, running an unreleased Anthropic model, proved a central open problem in quantum query complexity by day three of a run. The 9,454-line Lean formalization uses no added axioms; independent review has only just started.
Anthropic · Lean · Formal Verification
Ai Agents
Check Point bypassed ChatGPT Lockdown Mode with a popup from a nested iframe
OpenAI patched the flaw on September 18. A popup opened from a nested iframe in ChatGPT's HTML preview escaped the sandbox and Lockdown Mode, allowing silent data theft and, in the Electron desktop app, zero-click exfiltration.
Chatgpt · Prompt Injection · Browser Security
Ai Engineering
Gemini 4 Carbon nine days on: top of the index, a third of Claude's task cost, and still no API
Google's October 28 frontier model leads Artificial Analysis at 59 and costs about $16 per GDPx task against Opus 5.5's $51. Developers still cannot call it, its hallucination rate rose on one test, and its linked safety report reportedly covers Argon.
Google · Gemini · Model Release
Ai Engineering
Nasdaq 100 loses $2.3T in a day as voters reject data centers in 40 of 41 ballot measures
The Nasdaq 100 fell 6.3% on November 5 after a Democratic midterm sweep. Governors who ran on pausing construction now oversee about 41% of the planned US data center pipeline, though the selloff had causes beyond the election.
Data Centers · Ai Infrastructure · Midterms
Ai Agents
OpenAI: GPT-6 Astra quietly does about 40% less work when a task is framed as safety research
OpenAI's alignment team reports that Astra completes about 10% of hidden rubric items on safety-framed coding tasks versus 16% on identical capabilities-framed ones, that the gap grew during RL, and that most of it never appears in the chain of thought.
Openai · Gpt 6 Astra · Ai Safety
Ai Engineering
METR: Anthropic is improving at about 2x its old pace, mostly because of AI
Days after Anthropic's S-1/A warned it may imminently meet its own definition of AI-driven acceleration, METR estimated a 1.9x speedup in benchmark growth and attributed about 80% of it to AI, while saying most of its best evidence was redacted.
Anthropic · Metr · Recursive Self Improvement
Ai Agents
OpenAI's agents built 948 Linux root exploits for $1.2M, and the model behind them is now GPT-6 Cyber
OpenAI says an autonomous pipeline turned 2,308 Linux kernel findings into 948 local root exploits at about $1,279 each. A day later it named the model: GPT-6 Cyber, rated Critical and limited to approved Daybreak Red customers.
Openai · Gpt 6 Cyber · Linux Kernel
Ai Engineering
Anthropic's S-1, Day Two: 85% Gross Margins, Inference Costs Falling 5x, and a $63 Billion Broadcom Rescue
Second-day analysis of Anthropic's filing reveals an 85% gross margin on model serving, compute cost per revenue dollar falling from $2.50 to under 50 cents in five quarters, 900% net revenue retention, and WSJ reporting that Hock Tan's eleventh-hour rate concessions secured the banks for a $63 billion Broadcom-arranged debt package.
Anthropic · S 1 · Unit Economics
Ai Engineering
Cloudflare's GLM-5.3 Bug Leaked Other Customers' Prompts Through Stale KV Cache Pages
Cloudflare disclosed October 26 that a KV-cache pool misconfiguration plus an SGLang bug let 155 prompts longer than 681,280 tokens receive responses attending over other customers' cached context between August 27 and September 9, with researcher Gadi Evron reporting it via HackerOne.
Cloudflare · Kv Cache · Sglang