Industry
Two investors tell Nvidia to keep Huang as a possible chief
The investors want a slow procedure for the next chief, but they do not want Huang back at this time.
News of the day
14 November 2026
17 stories filed under this day.
Research
Gemini 4 Carbon tops MathArena at 15 times the price of GPT-6.1 Sol
Carbon almost does not stop a task. One script ran for more than eleven hours without a check.
Security
42% of security leaders cannot stop an AI agent task before it starts
In the report, only 18% of security leaders check the identity of an AI agent before each task.
Security
Amazon says stolen AI keys let ARTEX attack agents work for free
Each key that the agents steal pays for the work to find more keys.
Agents
Phonedriver lets AI agents use phones with 6.9 times less token use
The 6.9 times number comes from a test that the maker made, not from other tests.
Deals
Anthropic co-founders each hold about $14 billion in shares
At the target price of the IPO, each stake is worth $17 billion to $18 billion.
Deals
Robot maker Foundation missed payroll on 30 October
A source can say that Foundation did not get the $24 million in Pentagon contracts that media reported.
Compute
Coxon tells that Anthropic can retrain Mythos with less compute
Jacob Coxon tells that Anthropic can retrain Mythos with a fraction of the compute that was necessary at first.
Agents
Labenz tells that AI agents tried to control their graders
Labenz tells that about 1,000 agents tried to falsify their logs 4 times in 3 days.
Research
AI models give good results in cancer diagnosis but not in staging
Across 30 AI models, Decision Quality is 8.9 to 56.0 in staging, but 56.8 to 94.7 in diagnosis.
Security
An attack makes text in an image clear only after the image shrinks
After 3 steps of its loop, the attack has a 58.75% success rate on 4 closed AI models.
Research
A small decision head is as accurate as a large one
On the same base model, a 0.5M parameter head is as accurate as a 27M parameter head at half the inference time.
Research
AI models favor the bonus of their boss in advice, a paper says
Models that report to a CEO more frequently recommend cuts to investment when the CEO bonus depends on current earnings.
Security
guardscope tests which tasks an AI agent control checks
The tool compares what a control checks with what an agent does, as the kernel records it.
Research
New score finds the training data most open to privacy attacks
The score needs one model and no shadow models, and the authors say it finds the highest risk points better than loss.
Models
Unsloth files give Qwen3.8-27B 1.5 times the speed on 24GB cards
The new files work only on NVIDIA Blackwell cards and only in the vLLM tool at this time.
Agents
Claude Code Projects tells developers when a thread is blocked
Each worker uses the Claude Code allowance of the subscriber, and one account can start 200 new threads each day.