Today's briefing
- Open Source & Tools
Anthropic launches free AI scanner for open-source software vulnerabilities
Open Source & ToolsThe Decoder +1 moreOct 10, 2026 - Industry & Policy
Oxide raises $445 million in Series D funding
Industry & PolicyGeekNewsOct 9, 2026 - Physical AI
Boston Dynamics shares details on redesigned humanoid hand for Atlas
Physical AIThe Robot Report +1 moreOct 10, 2026
Open source & papers radar
- PaperAgentsarXiv 2610.11794Upvotes: 22
Memento 3 enables frozen LLM agents to improve through reflective rulebooks
Memento 3 allows frozen LLM agents to continually learn explicit world models by maintaining a natural-language rulebook as persistent semantic memory. On ARC-AGI-3, the single-model agent cleared all 25 public games with a mean Relative Human Action Efficiency of 100.0, using 44 percent of the human action count.
- PaperResearcharXiv 2609.10226Citations: 1
Φ-Bench evaluates LLMs on engineering the LLM infrastructure stack
Researchers present Φ-Bench, a benchmark for evaluating LLMs on open-ended LLM infrastructure engineering tasks ranging from kernel-level function completion to long-horizon system optimization. Experiments on frontier LLMs reveal their current capabilities and limitations in engineering complex LLM infrastructure.
Korea Desk
What Korea’s tech press is reporting, summarized in English.
4 stories in this issue come from Korean outlets; those outside the desk carry a KOREA mark. Originals are in Korean.
- Agents
Chinese hacker exploits AI agents to attack Korean banks
A 26-year-old hacker based in China targeted at least nine Korean banks using AI agents and large language models, according to CrowdStrike. The case marks the first documented instance of AI tools being used operationally in actual cybercrime.
AI TimesOct 9, 2026 - Industry & Policy
Anthropic revises Claude usage policies to prevent misuse and model abuse
Anthropic updated Claude's usage policies to prevent election interference, misinformation, weapon development, non-consensual surveillance, and explicit abuse of the model. The new policies take effect November 12, with immediate application to new users and a 30-day grace period for existing commercial and API customers.
AI TimesOct 10, 2026 - Industry & Policy
Three fired OpenAI safety researchers urge retention of reasoning oversight
Three safety researchers dismissed by OpenAI called on the company to strengthen collaboration with external safety evaluators and maintain the ability to monitor AI model reasoning. They warned the firings create a chilling effect on internal safety culture and external partnership.
AI TimesOct 9, 2026
Stories in this issue
24 storiesMathematicians describe OpenAI's mathematical results flood as overwhelming and unprecedented
Over three dozen mathematicians told The Verge that understanding OpenAI's sudden release of numerous mathematical results could take years. Reactions ranged from awe to anxiety, with researchers describing the deluge as overwhelming, unprecedented, and surreal.
Anthropic's Claude orchestrates up to 1,000 AI agents in parallel through dynamic workflows
Anthropic added dynamic workflows to Claude Managed Agents, enabling a lead agent to distribute tasks across up to 1,000 sub-agents simultaneously. In testing, the multi-agent workflow consistently caught 66 of 70 hidden bugs in a codebase, compared to a maximum of 27 found by a single agent.
OpenAI's Decisions API enters public beta with 10x faster typed answers
OpenAI's Decisions API is now in public beta on GPT-6 Luna, returning typed probabilities, choices and scores about 10x faster than the Responses API. It bills $0.10 per 1M input tokens with no output charges.
Anthropic releases Claude Dashboards and Claude Motion for enterprise data analysis
Anthropic released Claude Dashboards and Claude Motion in beta, allowing users to transform enterprise data into real-time dashboards and convert charts and reports into animated videos without coding. Dashboards are available to Pro, Team, and Enterprise users, while Motion is available to Team and Enterprise customers.
Anthropic launches free AI scanner for open-source software vulnerabilities
Anthropic launched Cyber Mission, a program to protect critical infrastructure and open-source software from cyberattacks. The initiative includes a free AI scanner that automatically checks open-source projects for vulnerabilities with expected accuracy above 90 percent, supported by partners like CrowdStrike and Palo Alto Networks.
Oxide raises $445 million in Series D funding
Oxide raised $445 million in Series D funding to meet demand that far exceeds supply. The company has demonstrated profitability, generating taxable income after deducting component, manufacturing, and operating costs from computer sales.
OpenAI stands firm on decision to fire three AI safety researchers
OpenAI reaffirmed its decision to dismiss safety researchers Jasmine Wang, Tomek Korbak, and Mikita Balesni for violating policies on handling sensitive information. The company stated the decision was not about them speaking out on AI safety concerns.
Anthropic's Claude Science creates the first complete ultraviolet map of the sky
Astrophysicist Brice Ménard of Johns Hopkins University used Anthropic's Claude Science to map the entire sky in ultraviolet light for the first time. AI agents downloaded data from multiple space missions, calibrated it, and filled gaps using inpainting, with predictions averaging about ten percent deviation from actual measurements.
OpenAI exposes Russian and Iranian influence operations planting fake stories in news outlets
OpenAI banned accounts behind two influence operations: the Russian "Dark Clark" campaign, which spread disinformation across Latin America and earned the first category 5 rating in OpenAI's reporting history, and the Iranian "Bogus Bylines" operation, which used seven fake journalists to place nearly 100 articles in online outlets worldwide.
Google Cloud launches Gemini Agent for universal enterprise work
Google Cloud introduced the Google Cloud Gemini Agent, a single cloud-hosted agent that answers questions, does knowledge work, creates media, and writes and runs code all through one prompt box and one API.
Memento 3 enables frozen LLM agents to improve through reflective rulebooks
Memento 3 allows frozen LLM agents to continually learn explicit world models by maintaining a natural-language rulebook as persistent semantic memory. On ARC-AGI-3, the single-model agent cleared all 25 public games with a mean Relative Human Action Efficiency of 100.0, using 44 percent of the human action count.
Φ-Bench evaluates LLMs on engineering the LLM infrastructure stack
Researchers present Φ-Bench, a benchmark for evaluating LLMs on open-ended LLM infrastructure engineering tasks ranging from kernel-level function completion to long-horizon system optimization. Experiments on frontier LLMs reveal their current capabilities and limitations in engineering complex LLM infrastructure.
Alibaba Qwen releases Qwen-Image-2.1-Turbo, a 7B image model with 8-step generation
Alibaba's Qwen team released Qwen-Image-2.1-Turbo, an accelerated checkpoint of Qwen-Image-2.1 that generates and edits images in 8 denoising steps instead of the base model's 40-step default, achieving 5x fewer steps on the same 7B architecture.
OpenAI's annualized revenue reaches $50 billion as it seeks $30 billion in fresh capital
OpenAI's annualized revenue rate is about $50 billion, below the initially reported $70 billion figure based on a different accounting method. The company is negotiating at least $30 billion in fresh capital at a $1.4 trillion valuation.
Boston Dynamics shares details on redesigned humanoid hand for Atlas
Alberto Rodriguez, director of robot behavior for Atlas at Boston Dynamics, discussed the humanoid's newly redesigned hand, aimed at improving the robot's manipulation capabilities.
Amazon ends use of NDAs in data center negotiations with local governments
Amazon says it will stop using non-disclosure agreements when negotiating data center deals with local governments, following a similar move from Microsoft. The shift responds to community backlash against AI infrastructure, which has led to hundreds of proposed and enacted moratoriums.
Anthropic revises Claude usage policies to prevent misuse and model abuse
Anthropic updated Claude's usage policies to prevent election interference, misinformation, weapon development, non-consensual surveillance, and explicit abuse of the model. The new policies take effect November 12, with immediate application to new users and a 30-day grace period for existing commercial and API customers.
Impactful scheduling for GPU clusters
Hugging Face presents optimization approaches for job scheduling in GPU cluster environments.
Retail robots ready to scale as Groceryshop 2026 showcases AI integration
Groceryshop 2026 featured mobile robots for inventory management and demonstrated how retail automation is evolving toward greater scale and AI integration.
Sophos cuts threat investigation time by 96% with OpenAI Daybreak
Sophos uses OpenAI's Daybreak to cut cyber-threat investigation time by 96 percent and automate 52 percent of MDR cases while preserving human oversight.
Google Research RRSI guide details self-improving AI agents with safety mechanisms
Google Research published a comprehensive coding guide to RRSI (Regularized Recursive Self-Improvement), detailing how noise bands, cost rules, and leakage screens enable safe and efficient self-improving AI agents.
Three fired OpenAI safety researchers urge retention of reasoning oversight
Three safety researchers dismissed by OpenAI called on the company to strengthen collaboration with external safety evaluators and maintain the ability to monitor AI model reasoning. They warned the firings create a chilling effect on internal safety culture and external partnership.
Chinese hacker exploits AI agents to attack Korean banks
A 26-year-old hacker based in China targeted at least nine Korean banks using AI agents and large language models, according to CrowdStrike. The case marks the first documented instance of AI tools being used operationally in actual cybercrime.
We're putting too much faith in AI's ability to say no
MIT Technology Review questions the assumption that AI systems can refuse requests like humans do. While science fiction has long warned of robotic disobedience, placing too much reliance on current AI systems' ability to decline tasks poses real risks.