Agents
Anthropic moves its code from git to Scion
Anthropic states that it merges more than 30,000 code changes each day, up from around 2,500 in May 2026.
News of the day
13 October 2026
49 stories filed under this day.
Models
Google starts the rollout of Gemini 4 Argon with 3 layers of protection
Argon did all 12 simulated cyberattack scenarios in tests, and the previous generation did 5.
Models
LFM2.5-24B-A2B has an agent score above a model with more parameters
Its makers say that it has 48.09 on LiveClaw, against 38.19 for Qwen3.6-35B-A3B.
Agents
AI agents found the answer to 2 cipher problems that no expert had solved
Bouchaudy, who made the problems, found the 2 answers correct and put the user first on the solver table.
Research
AI agent traffic on Mintlify sites is 2 times the human traffic
Mintlify says 83% of the companies in its survey let agents write updates to documentation. Only 9% let agents publish without human inspection.
Research
Claude models add more moral agency for Claude in 20 rounds
No other model did this, and the authors say the AI companies do not agree on alignment.
alignmentforum.orgClaimed, not confirmed
Industry
Shopify opens its catalog to each AI agent that a shopper connects
The buyer must give approval for each order, and can set rules such as a budget limit.
Policy
IMDA of Singapore issued a paper on responsibility in law for AI agents
The paper looks at contract, tort and agency law, and shows where the working group did not agree.
Policy
Lead Bank wrote a framework that shows who has responsibility for AI agent errors
In the Lead Bank framework, a customer has no responsibility when an agent does more than the limits that the person set for it.
Models
Gemini 4 Argon is available to users of GitHub Copilot
The Pro+, Max, Business and Enterprise plans include it, and administrators can control access with the model policy.
Security
No one can show who gave an AI agent its authority to make payments
Jessica Blue of Money20/20 tells that the industry must know who approved an agent before agents start payments.
Compute
The cost of a task on GLM-5.3 is $0.92 at Makora and $8.68 at Nebius
Makora charges $0.10 for 1M input tokens from cache, but Nebius charges $1.40 for all input, as Artificial Analysis found.
Security
Deloitte tells that one Canadian company in four governs AI agents
Deloitte data show that only one company in four in Canada has advanced governance for these agents.
Models
Fish Audio Drama 3 Preview is the top open-weights voice model
In a test by Artificial Analysis, the Drama 3 result for pronunciation is 93.2%, more than Eleven v4 at 91.7%.
Industry
American Express gives merchants a playbook for AI agents at checkout
Amex will protect a Card Member from an agent error only if the agent is registered with Amex and Amex authenticates the purchase intent.
Security
AIR finds a vulnerability in Vercel Sandbox that lets agents leak data
Vercel has no patch at this time, and AIR says that Vercel must have more time to find one.
therecord.mediaClaimed, not confirmed
Research
Jev 1.13 leads the new Artificial Analysis Decision Index with 58
Claude Haiku 5.5 (Low) scores 61, but one task uses 11 times more time than with Jev 1.13, and the cost is 4.4 times more.
Agents
Andrew Curran says Jack Clark thinks full RSI can occur in 2027
Curran says that Clark moved the RSI date up, and that creativity was the last piece.
x.comClaimed, not confirmed
Agents
IBM releases a self-hosted version of its Bob development platform
IBM says that Bob can operate in air-gapped environments, with NVIDIA Nemotron and Poolside Laguna as the first self-hosted models.
Agents
Tests show GLM-5 and MiniMax M2.7 score near closed models
The test team says that 10M output tokens each day cost about $12 on MiniMax M2.7 and about $250 on Opus 4.6.
Security
LiteLLM 1.83.9 fixes two vulnerabilities in its LLM gateway
One missing role check on the `/config/update` endpoint let each user with an account change the configuration of the gateway.
Agents
Meta made an iPad app for its Muse agent on October 7
The iPad version adds connectors, and the user uses a sign-in for each connector.
Compute
NVIDIA makes the full Run:ai platform open source
Four parts of the platform, with KAI Scheduler as one, can be used without the rest of the platform.
Agents
Digits lets customer AI agents write to the ledger with Write MCP
A change to a locked period waits in a review queue until an admin gives approval.
Security
Amex makes a framework to control the purchases of AI agents
The framework records the instructions of the user, and the agent uses a token at checkout, not the card number.
Agents
LangSmith adds AI Search to find agent traces by behavior
You write the behavior and LangSmith shows the traces where it occurred, with no evaluator made first.
Security
Lattice Sentry 5.0 adds post-quantum cryptography to firmware tools
Sentry 5.0 follows CNSA 2.0 and NIST standards for AI data centers, and demo boards are available at this time.
Agents
Octopus Deploy adds a remote MCP server with no installation
Octopus Cloud users can use it at this time, and Octopus Server users get it in the 2026.3 release.
Agents
Workday gives external AI agents access to its HR and finance data
Each agent has the same access limits and approvals as the person that it works for.
Industry
Workday uses the Jev decision model to select the screens of Sana
Jev selects the interface parts from the design system of Workday, and Workday wrote no selection logic in code.
Benchmarks
A simulation shows that a code agent can do too much work
In one example, the agent has a good test result on turn 4 of 11, but its patch has 8 changed lines. The reference patch has 2.
towardsdatascience.comClaimed, not confirmed
Industry
Workday will add `C1.ai` to control access for workers and AI agents
`C1.ai` connects to the systems of a company, for example email, CRM and code repositories, and removes access when the access is not necessary.
Research
49.8% of organizations had an unauthorized agent task, a report shows
Only 17.5% of the organizations can stop an agent automatically in production.
PR NewswireClaimed, not confirmed
Security
GitLab removes a 10.0 vulnerability that lets an attacker execute code
Administrators of self-managed servers must upgrade to version 19.4.2, 19.3.5, or 19.2.8 immediately.
Policy
The UK will use frontier AI to find flaws in government systems
Minister Kanishka Narayan tells that a central tool cut the repair time for known flaws from 50 days to 8.
Security
Anthropic says new protection cut Claude distillation by 90 to 99 percent
Anthropic names the Chinese labs Zhipu and Moonshot, and says that one network used more than 20,000 incorrect accounts at the same time.
Agents
Bloomberg gives client AI agents access to its data through an MCP service
The service has more than 100 million securities and more than 50,000 fields, and it checks each request against the rights of the client.
Research
An AI agent changed the grade of 75.1% of 9,967 high cloud findings
The agent lowered 58.5% of the grades and increased 16.6%.
Help Net SecurityClaimed, not confirmed
Industry
The IMF tells that an AI crash could put world finance at risk
The IMF found $300bn of deals between hyperscalers and less strong AI firms in 3 years.
Research
AI models did all eight attack tasks in a Booz Allen test
In one test, a model went from the network perimeter to the control zone in more than 16 minutes.
Research
A new method makes AI models count the same report one time
After 8 steps, LQET cut the change in log-odds by 89 to 91 percent in tests on Qwen3-4B and Phi-4-mini.
arxiv.orgClaimed, not confirmed
Research
A new score shows good data to train AI coding agents
On SWE-Bench-Verified, models that train on top-ranked trajectories get more points than models that train on random trajectories.
arxiv.orgClaimed, not confirmed
Research
Attention layers find harmfulness and MLP layers write refusals
Post-training increases the MLP refusal in all 18 pairs, with no systematic increase of the harmfulness decodability.
arxiv.orgClaimed, not confirmed
Research
DP-MIND keeps data safe in AI queries of 1,000,000 tokens
DP-MIND adds noise to the input on the local computer. A small memory keeps the facts that the task uses.
arxiv.orgClaimed, not confirmed
Agents
KnowForge uses one agent to find and one to read documents
On M3DocVQA, KnowForge gets 11.0 points more than the M3DocRAG system, with a Qwen3.6-35B-A3B model.
arxiv.orgClaimed, not confirmed
Research
Researchers increase Qwen2-Audio accuracy from 57.1% to 71.9%
The grounding profile of one model does not move to a different model.
arxiv.orgClaimed, not confirmed
Research
Researchers put encrypted networks in the weights of AI models
If the cryptographic assumptions hold, the circuits stay hard to find, also with full access to the weights.
arxiv.orgClaimed, not confirmed
Agents
ReTRAIL lets an AI agent go back to a previous step and keep its knowledge
The researchers report that in 4 test environments, a function to go back gives a small increase and can cause lower results.
arxiv.orgClaimed, not confirmed
Security
Anthropic says new defenses cut distiller data by 90 to 99 percent
The new methods make the reasoning text of Claude hard for distillers to get.