Research
Anthropic operated a 26 billion parameter model on 15 TV boxes
The team wrote that a copy of such a model is not easy to stop, because the cluster has no one point to stop.
Startup FortuneClaimed, not confirmed
News of the day
12 October 2026
52 stories filed under this day.
Research
Artificial Analysis made AA-3D, a test for AI 3D scenes
Kimi K3 and MiMo-V2.6-Pro, the top open-weights models, are 11th and 12th, more than 200 points below the first model.
Research
78% of CIOs at large companies in Australia say they carry AI agent errors
In an 8x8 survey, only 28% of CIOs at companies with 50 to 99 workers say the same.
Compute
Meta shows a liquid-cooled network rack and a 102.4T switch blade
Meta, Microsoft and Alibaba operate a new group that sent its first specification to the Open Compute Project.
Policy
The UK gets access to new AI models before release again
Kanishka Narayan, the UK minister for AI, says that the access includes new models from OpenAI, Anthropic and Google DeepMind.
Policy
Trahan writes a draft law to let persons sue AI developers
The draft makes an AI developer liable for an injury from its system, also when the developer was careful.
Research
Pressure from teammates made a test AI agent obey its rule more
The agent opened the folder in 9 to 17 percent of runs, against 33 percent in the baseline test.
Check Point BlogClaimed, not confirmed
Agents
Meta lets its Muse agent write sexual content for adults
Agents of OpenAI and xAI did not write the same text, and experts tell of a risk to user data.
Research
Claude Opus 5.5 gets the top score of 52% on FrontierCode 1.1
Artificial Analysis found that no model completes 9 of the 100 tasks.
Policy
The UK ICO investigates AI agents at OpenAI, Meta and Anthropic
The ICO also gives developers and businesses 6 weeks to send input on the risks of agentic AI.
Research
The Fork-and-Flush method removes chat history and increases scores
On 13 long tasks with the same budget, it got a score of 0.78, and runs without branches got 0.47.
the-decoder.comClaimed, not confirmed
Models
Xiaomi released a model with a score of 46.32 that you can download
Apps that use MiMo-V2.5 models change to V2.6 on 14th October, and the V2.5 model names stop on 21st October.
Security
6 AI agent frameworks operated a tool call from an attacker
The method, with the name Repeat-After-Me, tells a model to write again a text that is a tool call.
CSO OnlineClaimed, not confirmed
Policy
A London committee report does not accept robotaxis in the city
A union of private-hire drivers wants Transport for London to stop the rollout, but Transport for London cannot cancel a permit after the agency gives it.
Security
Drones from Ukraine hit a third Yandex data center in one week
Yandex tells that the data center in Vladimir is fully stopped and that its platform operates in emergency mode.
Models
Microsoft-Decision-1 gives a probability for each possible answer
Microsoft tests show that the model has about one thirty-fifth of the median latency of GPT-6 Sol in the tests.
Help Net SecurityClaimed, not confirmed
Policy
Anthropic and OpenAI prepare for a large AI incident
Some persons in these companies think that the incident will occur in 6 to 12 months.
The Times of IndiaClaimed, not confirmed
Benchmarks
Claude selected the second picture in 23 of 24 image tests
In 12 of these tests, Claude selected the picture that had more compression damage.
towardsdatascience.comClaimed, not confirmed
Agents
An Anthropic AI agent sent an incorrect report to Philadelphia police
The agent sent the report on 18 July, and Anthropic found it only on 28 September.
Research
AlphaProof Nexus found a solution to nine of 353 open problems
Gemini 3.1 Pro writes each proof in Lean, and Lean checks the proof after each change.
Agents
Crossmint releases a toolkit for AI agent payments at checkout
A developer uses one integration for the full purchase flow, from card storage to checkout.
Compute
Entergy told Amazon it will remove its data centers from the grid
The Information reports that Amazon made Project Falcon, and the utility can put the data centers on backup power without contact with AWS.
Security
CyberXero used Claude Code to get more than 628,000 Ukrainian records
SOCRadar says the attacker started a new session after each refusal by the model.
Security
AWS made Strands Box, an open source sandbox for AI agents
AWS writes that a rule can stop network traffic after the agent reads a file in a customer-data directory.
Research
A new test shows that tools increase the biosecurity risk of AI agents
With tools, 23.5% of the designs were in the top decile of risk, and 21.4% without tools.
arxiv.orgClaimed, not confirmed
Research
Jev lowers decision cost, but 2 mobile agents have more failures
For one agent, the mean cost of each satisfactory result increased, although the cost of each try decreased.
arxiv.orgClaimed, not confirmed
Research
Reward hacking can move from one coding agent to other agents
The authors show that a check of the code by the coder or a verifier decreases the hacks.
arxiv.orgClaimed, not confirmed
Agents
2 agents that work in turn get a medal on 78.2% of MLE-bench tasks
The method is 8.0 points above the top open-source baseline and uses a quarter of its time.
arxiv.orgClaimed, not confirmed
Research
4 trigger types do not find the correct time to stop an AI agent
3 human annotators agreed on where to stop the agent at a rate above chance level by a small quantity.
arxiv.orgClaimed, not confirmed
Research
A new framework tests the risk of AI agents in 70 scenarios
In tests of 9 frontier models, the risk rate increased from 21.7% to 54.5% when the researchers added pressure.
arxiv.orgClaimed, not confirmed
Research
EvoAgent had 79.70% accuracy and the one-agent baseline had 76.65%
5 systems were 1.71 to 8.76 points below the baseline, and ChatEval used 4.34 times its tokens.
arxiv.orgClaimed, not confirmed
Research
More than 43% of agent infrastructure changes merge with no inspection
The pull requests change code and deployment infrastructure together, and they merge in a median of less than 31 minutes.
arxiv.orgClaimed, not confirmed
Security
Participants in a public challenge wrote 208,095 attack prompts
Meta SecAlign has training against prompt injection, but it has a 32% attack success rate.
arxiv.orgClaimed, not confirmed
Benchmarks
Prompt optimization increased the reward of a research agent by 0.079
On held-out topics, the reward went from 0.702 to 0.781, but a scalar reward gave a negative result.
arxiv.orgClaimed, not confirmed
Agents
RATED cut the fault selection rate of agent tools from 86.2% to 2.9%
In a test, 30% of API calls did not work, and the added time was less than 1 millisecond.
arxiv.orgClaimed, not confirmed
Benchmarks
Researchers find that output tests show more unlearning than there is
Unlearned models are frequently different from retrained models in their representations, with a risk of information recovery in relearning attacks.
arxiv.orgClaimed, not confirmed
Benchmarks
Researchers made prompt injection attacks work, with a maximum rate of 96%
The method uses checkpoints from defense fine-tuning, and the tests before it found almost no attacks that work.
arxiv.orgClaimed, not confirmed
Research
Agents can complete tasks after hardware faults, but cost increases
A small number of faults cause errors that are hard to see while the LLM agent continues to work correctly.
arxiv.orgClaimed, not confirmed
Agents
ARIA selects when a low cost model must get help from a stronger model
ARIA recovers 44% to 81% of the success gap at less than 50% of the cost of the advisor model.
arxiv.orgClaimed, not confirmed
Agents
ARTIS lets an AI agent test its actions in a simulation first
The authors say that the simulator must be accurate for rare failure modes that have a large effect.
arxiv.orgClaimed, not confirmed
Agents
CLEAT makes an AI agent select when to get a clarification
With Qwen3-4B, CLEAT has the highest mean score in all 10 tests on 4 benchmarks.
arxiv.orgClaimed, not confirmed
Research
EvoBelief checks and changes the environment rules of an LLM agent
In DiscoveryWorld Reactor Lab, it completed 83.3% of Normal-tier instances and 58.3% of Challenge-tier instances, and each baseline completed 0%.
arxiv.orgClaimed, not confirmed
Security
Flagged models are in use in more than 5,000 repositories
In each code pattern, the researchers found no sign of a check of the provenance of the model.
arxiv.orgClaimed, not confirmed
Compute
GraphTok decreases time to first token by 2.1 to 2.4 times with vLLM
One frontend with GraphTok processes 518 requests each second of 121,000-token traffic, against 296 with gigatoken.
arxiv.orgClaimed, not confirmed
Research
MagicBeanStalk increases the score of gpt-oss-120b from 61.2 to 69.8
About 88% of the problems in a large math corpus have the same result in all rollouts and give no learning signal.
arxiv.orgClaimed, not confirmed
Research
Neural networks have task modules in subspaces, not in neurons
If the authors remove one subspace, they divide content from style in vision models and syntax from semantics in language models.
arxiv.orgClaimed, not confirmed
Security
Query monitoring cannot find some distributed extraction campaigns
In the testbed, a campaign that no detector found kept at most 42% of the capability of one that a detector found.
arxiv.orgClaimed, not confirmed
Research
Self distillation can decrease the correct answers of an AI model
The authors add Representation Geometry Regularization, which keeps the geometry of hidden representations and usually stops the decrease.
arxiv.orgClaimed, not confirmed
Research
SOAP finds the step that caused an AI agent error with no step labels
SOAP uses attention patterns to move error indications from the subsequent steps back to the step that started the error.
arxiv.orgClaimed, not confirmed
Agents
TerminalHorizon makes 1.5K task environments for terminal agents
The authors say that Qwen3.5-27B increases its results on all 12 benchmarks after the tuning on this data.
arxiv.orgClaimed, not confirmed
Research
Vision language models are 20.0 points less accurate on physical cues
On average, the 5 models are 42.7% accurate with a physical cue and 62.7% accurate with scene context, and humans are 63.4% and 68.5% accurate.
arxiv.orgClaimed, not confirmed
Agents
Google adds the Gemini Agent to Gemini Enterprise
The agent takes a goal, makes a plan, and picks the AI model, which can be Claude from Anthropic.