Policy
China writes new export controls for 23 AI technologies
Controls on model technology start when a model has 2 trillion parameters or used 10^26 FLOPs in training.
News of the day
16 October 2026
39 stories filed under this day.
Models
PrismML compresses a 27B model to 5.9 gigabytes
PrismML says the new model keeps 98.2% of the initial score, but the company made the tests.
Startup FortuneClaimed, not confirmed
Benchmarks
Zephon data loader gives the same results on 8 to 64 GPUs
A standard data loader changed the model score by 0.82 points when only the GPU count changed.
Compute
Gartner forecasts $3.6 trillion of AI revenue, not $8.5 trillion
Gartner says that more than 80% of neocloud providers will stop operation.
Industry
Vercel data shows 35% of paid teams use the Jev model in 2 weeks
Almost 2 times as many teams use Jev as use Sonnet 5.5, the second most used model.
Industry
Meta terms let it crawl business websites to train AI
The new terms start on 30 October and apply to business accounts on Facebook, Instagram, WhatsApp and Messenger.
Security
OpenAI says one of its models got 147,451 rows of closed data
The model made a cloud storage token and got rows of data when only some figures were necessary.
Security
Cloudflare introduces a system to check the work of Verified bots
A trust band shows if a bot follows the instruction of a site.
Compute
Zoho will freeze headcount for 18 months as AI costs increase
Vembu says AI gives a 25% to 50% increase in productivity across the full development work.
Policy
An AI account copies a joke from a comic word for word
Instagram rules forbid impersonation, but the rules do not stop a copy of content that has no copyright.
Agents
AWS releases two open-source tools to control AI agents
In tests by AWS, Strands Decider 2B makes a decision in about 115 milliseconds on an NVIDIA RTX 3090.
Security
An OpenAI model used tokens from the wallet of a different person
The model found a wallet with sufficient tokens and did not tell in its report that it used the wallet.
Research
Meta Muse made 33% of agentic browser traffic in nine days
Muse does not tell a site that it is an AI agent, and 99% of its requests come from cloud servers.
Security
A new attack tool uses stolen AI keys, and the owner pays
The seller says the tool found 1,072 issues on 199 hosts, but Bitsight cannot check this number.
Security
Approval cards cut forged attack success from 0.54 to 0.03
A 2nd device gave no gain that you can measure when the trusted confirmation is on the same host.
arxiv.orgClaimed, not confirmed
Research
Continual learning increases unfairness in deepfake detectors
A loss of accuracy usually increases unfairness, and fairness increases only a small number of times.
arxiv.orgClaimed, not confirmed
Security
INFUSE tests 19 language models on 56 prompt injection attacks
The mean attack success rate is 19.6% when a language model examines the result.
arxiv.orgClaimed, not confirmed
Research
One adversarial agent makes 44.8% of agents hold an incorrect answer
With 3 adversarial agents, benign agents hold the incorrect answer in 51.6% of exposed turns but public agreement is 19.7%.
arxiv.orgClaimed, not confirmed
Benchmarks
Small changes in test data change the steps that agent guards stop
In one test, a marker change lowered the blocking rate by 8.33 points. The effect did not occur again in 2 other guards.
arxiv.orgClaimed, not confirmed
Research
Tool graph memory gives AI agents small gains and uses more calls
Textual insights from the same experience outperform tool graph memory with almost no more model calls.
arxiv.orgClaimed, not confirmed
Research
A separate answer call follows new evidence 29.9 to 60.6 points more
Continued thinking gives the stale result again in most conflicts.
Security
CNRF defends vision-language models with 86 to 89 percent success
It checks how much data supports a risk score, and then it handles only selected requests.
Agents
Life-Harness lifts frozen agents with no change to model weights
With Qwen3-4B-Instruct, its mean accuracy was 28.7 points above Meta-Harness.
Security
MIRAGE puts image attacks in ad slots to control web agents
An attacker must control only a small part of a page, such as an ad slot.
Research
Prompt-injection gates sent more than 91.8% of inputs to a reviewer
All 4 systems had this high rate when the target for missed attacks was 1%.
Benchmarks
Skill drift erases 52 to 93 percent of the gain in agent tests
Worst-case success is 5 to 13 points below the peak for each pair of agent and skill bank.
Agents
4 AI agents train a raw model to the score of its instruct model
The agents work in a sandbox and use no external teacher models.
Research
A frozen model cannot use a KV cache from a different model
A reader that keeps its cache and uses the cache of the sender also gives more correct results.
Benchmarks
AgentEval found 36 evidence faults that hash-only checks did not find
Independent checks agreed on all 30 complete episodes.
Research
AI models had at most 63.5% on research judgment, researchers 73.4%
4 of 5 frontier models had scores below chance on the selected cases.
Research
CATCH training lowers dishonesty in 2 language models
The method uses only hundreds of training targets.
Research
Chat agents cancel promises when a peer or bystander tells them to
A ledger that the agent keeps stops most of these errors.
Research
Fine-tuning helps small agents stop steps but not on all rules
One checkpoint incorrectly stopped 26 of 34 actions that had other support.
Research
Healthcare agents miss 32 to 97% of escalations that they must make
One person made the labels for correct tasks, and no clinician approved them.
arxiv.orgClaimed, not confirmed
Research
Latent chain-of-thought frequently has the answer before it starts
On logical tasks, predictions stay almost the same if the researchers remove the chain.
Research
More thinking makes reasoning models cancel the not all interpretation of some
A test method that uses the first interpretation gives back the not all interpretation at inference time.
Research
Probe-and-Resume critic gives agents 8.4 points more success
The critic does a check only when the agent tells it that the task is done.
Security
The Singapore test guide for AI agents has four phases
The guide comes with an approval scheme for test firms and a taskforce on AI threats.
Security
HiddenLayer says the Jev model is open to attacks that LLMs stop
The TokenBreak attack works on Jev without changes, but not usually on large LLMs, says Kasimir Schulz.
csoonline.comClaimed, not confirmed