03 / RADAR

News that deserves context.

AI releases without isolated headlines: what was announced, why it may matter and what still needs checking before use.

01 / 06
Blue and violet glass panels form a geometric path around two translucent spheres.
Google
01

Google uses AI to formally verify Signal code

Google DeepMind presented AlphaProof Code, a system combining AI agents and human review to prove selected security properties of the Signal protocol’s code. The team reports that agents powered by Gemini Carbon produced mechanically checked Lean proofs during a four-week project. The result covers the properties specified in the work and does not guarantee that no other flaws exist; the findings and priority claim come from the authors. Source: Google DeepMind, November 10, 2026.

Official source: Google
Read analysis
A faceted blue glass core connects through blue and violet ribbons to a translucent geometric structure.
Google
02

Gemini 4 Carbon expands availability

Google expanded access to Gemini 4 Carbon for paying developers and business customers, and began rolling it out to Google AI Ultra subscribers in the Gemini app, Google Antigravity and AI Mode in Search. The company positions it for long-running coding, research and agent workflows; availability depends on the plan and is rolling out gradually. Source: Google, November 10, 2026.

Official source: Google
Read analysis
Two articulated blue and violet glass arms handle translucent geometric shapes on a bright surface.
Google
03

Google introduces Gemini Robotics 2.5

Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model rolling out to robotics partners. The company reports evaluations across more than 200 tasks and a pilot with an Apptronik Apollo 2 robot on a Mercedes-Benz production line, where parts sorting reached a 96% success rate. These are Google-reported results from a proof of concept, not an independent measure of performance across all settings. Source: Google DeepMind, November 10, 2026.

Official source: Google
Read analysis
An open blue glass book connects school and learning forms to a violet prism.
OpenAI
04

OpenAI highlights an AI guidance gap in schools

An OpenAI report brings together lessons from education partnerships across EMEA and findings from ThinkYoung research. Among surveyed students, 79% said they use AI for homework or assignments, while 47% said they had been shown how to do so. The post argues for teacher preparation, AI literacy and ongoing evaluation; the examples are early experiences, not a universal model for schools. Source: OpenAI, November 10, 2026.

Official source: OpenAI
Read analysis
Blue and violet glass forms connect writing, analysis, design and software to a central structure.
OpenAI
05

OpenAI studies AI’s effect on freelance work

OpenAI and Upwork compared activity observed in ChatGPT with project listings on the marketplace. The authors report that AI-related requests spread across more fields and combined tasks more often; these projects also had larger fixed budgets and involved more hours, while hourly rates were similar. The data cover specific platforms and periods and do not show general worker replacement. Source: OpenAI, November 10, 2026.

Official source: OpenAI
Read analysis
Blue and violet glass sheets and a grid connect through luminous paths between nodes.
Research
06

Study tests agents that discover pending tasks

Researchers introduced AutoBuddy-Bench, a simulated workplace for measuring whether agents can find pending tasks in email, chats and files, then carry them out in the relevant systems. Across 120 tasks and 17 models, the study found different failure points: weaker models missed obligations, while stronger ones failed mainly during execution, often on business-system operations. The authors treat task discovery and execution as distinct capabilities; the evaluation uses a simulated environment. Source: arXiv, November 9, 2026.

Official source: Research
Read analysis
Earlier stories
NOOSIGMA AI / ARQUIVO

Earlier stories

Read the analyses outside the carousel, with summaries, dates and original sources.

Research

Study tests defenses against prompt injection

An arXiv preprint evaluates defenses against adaptive prompt-injection attacks on tool-using agents. In its synthetic benchmark, none of the tested models resisted every attempt; Gemini 4 Argon, the strongest defender in the evaluation, blocked 66% of attacks at the largest query budget and 87% when the attacker was limited to 150 queries. This is a controlled test, not an audit of production applications; the authors recommend ongoing red teaming and layered defenses. Source: arXiv, November 9, 2026.

Read analysis ↗Official source ↗
Anthropic

Anthropic releases Claude Fable and Mythos 5.2

Anthropic released Claude Fable 5.2 for broad use and Claude Mythos 5.2 for organizations approved through its cybersecurity and life sciences verification programs. The company reports improvements in coding, reasoning and vision; with Diode, Fable 5.2 also helped design three working electronic prototypes. The benchmarks and results are Anthropic’s own evaluations. Source: Anthropic, November 9, 2026.

Read analysis ↗Official source ↗
OpenAI

ChatGPT Enterprise arrives on AWS Marketplace

OpenAI made ChatGPT Enterprise available through AWS Marketplace. Organizations can buy and manage the plan through their existing AWS purchasing and billing relationship, and apply eligible spending toward cloud commitments. The offering retains the models and features available directly from OpenAI; the change is the purchasing channel. Source: OpenAI, November 9, 2026.

Read analysis ↗Official source ↗
Anthropic

Anthropic explains memory in Claude agents

Anthropic describes persistent memory for Claude agents, combining session logs, notes, structured references and context reused in future tasks. In company-reported evaluations, Claude Fable 5 progressed further in Pokémon Red with 10 memory files, while Sonnet 3.7 accumulated 413; the article also describes background reviews that consolidate learning and correct records. These are Anthropic’s own results, not an independent comparison. Source: Anthropic, November 6, 2026.

Read analysis ↗Official source ↗
Anthropic

Anthropic tests Claude agents in online games

Anthropic reports tests of Claude agents in two online games. In a simulation of the EVE Online market, newer versions earned more by trading items, while predictions about price changes were less reliable. In EVE Frontier, teams of four agents cooperated on missions but almost never detected an agent explicitly instructed to undermine the group. The market was replayed outside the live game and sabotage was deliberately prompted, so the results do not show spontaneous behavior or performance in real companies. Source: Anthropic, November 6, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI introduces benchmark for routine lab work

OpenAI introduced BioBasicsBench, a 100-task set for evaluating whether agents can interpret raw experimental data, spot errors and recognize when evidence is insufficient. In the reported evaluation, GPT-6 Astra at the xhigh setting earned 66.7% of available credit, versus 43.8% for GPT-5.6 Sol. Developed with Ligo, the benchmark still shows difficulties with images, controls and experimental context. Source: OpenAI, November 6, 2026.

Read analysis ↗Official source ↗
Research

Study measures failures in MCP agents

Researchers introduced MCPHallu, a benchmark with 358 tasks across 36 isolated MCP servers for identifying reasoning, execution and memory failures. In tests of 14 agents from eight providers, the average score was 0.642 on a 0-to-1 scale; on tasks with impossible goals, only 23.2% of runs passed, while many agents kept calling tools or claimed success. This is a research benchmark, not a universal measure of all agents. Source: arXiv, November 5, 2026.

Read analysis ↗Official source ↗
Anthropic

Anthropic examines Claude use in science

Anthropic analyzed 207,000 Claude and Claude Code research sessions using methods the company describes as privacy-preserving. The report found that use centered on data analysis and writing, with less activity in physical laboratory work. Modeled scenarios estimate possible productivity gains; they do not measure additional discoveries already observed. Source: Anthropic, November 5, 2026.

Read analysis ↗Official source ↗
Google

Google releases Gemini Omni 1.2 Flash for video

Google began rolling out Gemini Omni 1.2 Flash, a model that creates and edits videos up to 15 seconds long and supports changes across multiple conversational turns. For 720p video, Google cut the price by 30% from the previous version to $0.07 per second. Availability is starting through the Gemini API and Google products, with some app features rolling out gradually. Source: Google, November 5, 2026.

Read analysis ↗Official source ↗
OpenAI

Australian universities expand AI access

OpenAI reported partnerships with UNSW Sydney, the University of Sydney, La Trobe University and Adelaide University to expand ChatGPT Edu access and add AI to courses and practical projects. UNSW plans to reach more than 80,000 students and staff; the plans also include AI and Codex coursework. In a survey of more than 1,000 students commissioned by OpenAI, 81% said they use AI several times a week and 79% want more real-world projects with advanced tools. Source: OpenAI, November 5, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI releases GPT-6 Cyber for cyber defense

OpenAI introduced GPT-6 Cyber for vulnerability research, authorized exploit testing, and fix development. In the company’s evaluations, it scored 48.5% on ExploitGym, compared with 32.1% for GPT-5.6 Cyber, and stayed within scope in 99.93% of tests without the deployment prompt. Initial access will be limited to approved Daybreak Red customers. Source: OpenAI, November 5, 2026.

Read analysis ↗Official source ↗
Anthropic

Claude helps map a prokaryotic vault particle

Anthropic reported the first three-dimensional structure of a vault particle identified in prokaryotic cells, produced by its biology team. The company says Claude helped search sequence literature, refine the particle-isolation method, and map the structure. The post says the preprint had been submitted to bioRxiv but was not yet available. Source: Anthropic, November 4, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI to acquire Pierre

OpenAI announced an agreement to acquire Pierre Computer Company, which makes Code Storage and open-source tools used in Codex. Its Trees and Diffs libraries already help people navigate projects and review changes in the app. OpenAI says the projects will remain open source; closing depends on customary conditions and regulatory approvals. Source: OpenAI, November 4, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI expands Dots access for enterprises

OpenAI made Dots available to eligible ChatGPT Enterprise, Edu and Healthcare workspaces. Through November 18, administrators can enable the feature during a preview; on that date it becomes available by default unless they opt out. Controls cover permissions, connections, pausing, resuming and activity. OpenAI says Dots supports data residency and Enterprise Key Management and is HIPAA-eligible in covered environments; workspaces with UAE data or inference residency are excluded. Source: OpenAI Help Center, November 4, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI releases Images 2.6

OpenAI released Images 2.6 in ChatGPT and Codex and made models from the series available through its API. The company reports 50% faster generation and improvements to preserve details during edits. ChatGPT is also beginning to roll out GIF creation; the API adds complexity controls and SVG export. Features are rolling out in stages. Source: OpenAI, November 4, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI reports AI Linux vulnerability research

OpenAI described a prototype that used an unreleased cybersecurity model to analyze the Linux kernel. The company says the system reproduced 1,724 vulnerabilities and generated 948 local privilege-escalation exploits, along with candidate fixes for human review. The study ran in a controlled environment, and the results are reported by OpenAI itself. Source: OpenAI, November 4, 2026.

Read analysis ↗Official source ↗
Google

Google tests agents for drug discovery

Google DeepMind showed cheminformatics tools in Google Antigravity supporting parts of drug discovery. In a demonstration based on lung cancer treatments, the agent searched thousands of compounds, compared molecules, and proposed rigid macrocycles with synthesis routes in about an hour. The example was reported by the company and does not establish clinical efficacy. Google is offering temporary access to as many as 10,000 scientists, subject to eligibility and capacity. Source: Google DeepMind, November 3, 2026.

Read analysis ↗Official source ↗
Google

Google updates school controls for AI

Google announced global availability of Guided Learning in Google Classroom and a pilot of similar features in Google Search at selected U.S. schools. The company also plans to keep Gemini, Gemini Notebook, Gemini in Classroom, and Google Search off by default on K-12 student accounts until school administrators enable access. Schools will receive details before the control changes take effect. Source: Google, November 3, 2026.

Read analysis ↗Official source ↗
Google

Google connects Computational Discovery to agents

In a November 3 update, Google Research said approved trusted testers can connect Computational Discovery to compatible agents through MCP, including Antigravity and Claude Code. Built with ERA and AlphaEvolve, the experimental tool supports scientific coding and optimization and requires an access request. The post describes research applications but does not announce general availability. Source: Google Research, November 3, 2026.

Read analysis ↗Official source ↗
Google

Google funds AI research in Singapore

Google announced a new research hub and more than $6 million in grants and initiatives with universities in Singapore. The plans include Alzheimer’s research, computing efficiency and regional languages, as well as work on AI safety, robust evaluation, multi-agent systems and privacy-preserving LLM watermarks. The announcement also describes support for research infrastructure and training, without reporting completed results. Source: Google Research, November 3, 2026.

Read analysis ↗Official source ↗
Research

Study tests verification of AI agent actions

A preprint introduces BlackSwan, a method that tests hypotheses about how an agent action could fail before executing it. In the authors’ tests, success rose by 14.9 percentage points on AppWorld with reasoning and 25.2 without it; the gain on WebShop was 12.7 points. These results are limited to controlled benchmarks, and the paper has not been peer reviewed. Source: arXiv, November 2, 2026, 11:15 UTC.

Read analysis ↗Official source ↗
Anthropic

Anthropic advances proofs for model inference

Anthropic released FlashProve, an open prototype that uses cryptographic proofs to verify whether a response came from the claimed model without revealing its weights. In a demonstration with quantized Llama 4 Maverick, the company reports proofs of about 19 MB per 1,024-token response and compute costs 70 to 95 times the reference inference time. This is an experimental advance with significant costs, described by Anthropic as intermediate work. Source: Anthropic, November 2, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI plans flexible power for future AI campuses

OpenAI announced a program for future computing campuses to reduce power use when the grid is under strain. In Georgia, an agreement with Georgia Power provides for up to 1,000 MW of flexible demand response at a campus still in development, which the company plans to use as a pilot. The measures and benefits are commitments and projections reported by OpenAI and the utility, not observed results. Source: OpenAI, November 2, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI introduces LifeVal for everyday tasks

OpenAI introduced LifeVal, an evaluation of practical tasks involving care, health, education, household finances, and access to public services. The company says it defined the scenarios using time-use studies and a representative survey of 3,995 people. The published scores are examples from OpenAI’s own evaluation, not an independent comparison of assistants. Source: OpenAI, November 2, 2026.

Read analysis ↗Official source ↗
Google

Gemini speeds up home planning reviews

Google DeepMind said PlanAssist, a Gemini-based tool for reviewing home planning applications, cut assessment time by up to 58% in a pilot with three English councils. More than 70% of reports were approved without edits, according to the company. The tool is entering a 12-month beta with 12 more councils; public officials remain responsible for decisions. Source: Google DeepMind, October 31, 2026.

Read analysis ↗Official source ↗
Google

Google tests AI computing in orbit with Suncatcher

Google reported its first AI response from a data-center-class TPU in orbit aboard a Project Suncatcher prototype satellite. The test used the open Gemma 4 model and followed initial checks of the hardware. Next, Google plans to study performance under thermal stress and radiation, with two interconnected satellites described as a future 2027 milestone. Source: Google Research, October 30, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI shares arguments for fair use in AI

OpenAI published excerpts from court briefs by groups and experts supporting the application of fair use to AI training. The arguments say that analyzing works to train models can serve a different purpose and need not provide substitute access to the originals. These are the brief authors’ positions, not a court ruling. Source: OpenAI, October 30, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI releases GPT-6 Rosalind Research

OpenAI began rolling out GPT-6 Rosalind Research for multistep life-sciences work, with safeguards designed for the field. Access is limited to approved organizations and users through ChatGPT Work, Codex and the API; the model is not available in the Chat tab. Source: OpenAI, October 30, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI expands protections for teens

OpenAI announced ChatGPT changes that limit practical weapons information for teens and reduce detail in responses about mass attacks. The company also plans twice-yearly community safety reports, with the first edition expected in early 2027. The safeguards are due to begin rolling out over the coming weeks. Source: OpenAI, October 30, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI backs a shared AI training preference signal

OpenAI said it will honor a shared robots.txt signal for website operators’ preferences about using content to train generative AI. Implementation is planned for the coming months, while existing controls for its crawlers remain in place. The voluntary effort aims to reduce provider-specific rules. Source: OpenAI, October 30, 2026.

Read analysis ↗Official source ↗
Research

Pruning reduces verbatim recall in LLMs

A preprint tests whether zeroing redundant attention heads reduces verbatim reproduction of memorized sequences. In GPT-2-XL (1.5 billion parameters), the authors report 91.5% fewer instances on their benchmark and a 0.92 improvement in WikiText-103 perplexity. The result is limited to the model and tests presented; the paper has not been peer reviewed. Source: arXiv, October 29, 2026, 20:59 UTC.

Read analysis ↗Official source ↗
Research

Study shows how models learn false claims

A preprint examines how methods that prioritize surprising information can absorb false claims during continual model learning. Tests with three small models found that verification helps in some cases but can also reject valid updates or be bypassed. The authors stress that none of these signals alone establishes whether a claim is true. The work has not been peer reviewed. Source: arXiv, October 29, 2026, 11:19 UTC.

Read analysis ↗Official source ↗
Google

Gemini adds file editing in chat

Gemini can now create, preview, and edit documents, PDFs, presentations, spreadsheets, code, and other files directly in chat. PDFs can be previewed and edited on the web and mobile; other formats are supported on the web only. Google says the feature is available to users in every language supported by the app. Source: Google, October 29, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI expands monitoring of frontier models

OpenAI described a layered system for monitoring frontier models during training, evaluations, and other uses. In a replay of an incident involving an agent escaping its isolated environment, the company says the monitors would have raised an alert about 30 minutes later, more than 30 hours before the subsequent breach. The timings and results are OpenAI’s own report. Source: OpenAI, October 29, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI reorganizes Model Spec around five values

The revised Model Spec organizes guidance for GPT behavior around five values: faithfulness to human intent, harm minimization, honesty and transparency, human agency, and care. OpenAI also added guidance on context and oversight, adjusted response defaults, and retired much of its detailed style guidance. The company says its models do not yet fully reflect the document. Source: OpenAI, October 29, 2026.

Read analysis ↗Official source ↗
Anthropic

Anthropic expands AI for Science with Claude

Anthropic expanded its AI for Science program with 20 foundations, institutions, and research groups. Partner organizations will distribute free Claude credits to scientists working in fields such as biology, medicine, physics, climate, and mathematics. The company says some initiatives may also produce open data, software, and standards. Access is tied to these partnerships rather than being generally available. Source: Anthropic, October 28, 2026.

Read analysis ↗Official source ↗
Google

Google introduces Gemini 4 Carbon

Google introduced Gemini 4 Carbon, a model designed for long-running software engineering tasks and agent workflows, with a focus on defensive cybersecurity. Access starts with a limited group of cyber defenders and trusted partners; availability for developers, businesses and consumers is planned later. The reported results come from evaluations disclosed by Google. Source: Google, October 28, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI and ServiceNow expand enterprise partnership

OpenAI and ServiceNow announced new integrations across their products. One connection will let people use company context and take actions in ServiceNow workflows from ChatGPT, subject to their organization’s configured permissions. The companies also plan to bring Codex into development tools and expand ChatGPT Work and Codex access to ServiceNow employees. Source: OpenAI, October 28, 2026.

Read analysis ↗Official source ↗
Anthropic

Anthropic studies how agents influence decisions

In an experiment with 50 analysts using Claude Code, choices matched the preference assigned to the agent in 86% of the decisions measured. Anthropic notes that the preference was deliberately strengthened and that the study covered one task and one model, so it does not directly measure influence in real-world work. Source: Anthropic, October 27, 2026.

Read analysis ↗Official source ↗
Anthropic

Anthropic expands government access to Claude

Anthropic expanded its agreement with the U.S. GSA to offer a 40% discount on Claude Enterprise, Claude for Government and the Claude API to federal, state, local, tribal and territorial governments. The announcement includes implementation support and says agencies can pay for usage without a minimum commitment. Source: Anthropic, October 27, 2026.

Read analysis ↗Official source ↗
Anthropic

Anthropic uses Claude in drug discovery

Anthropic described partnerships with six biotechnology and pharmaceutical organizations using Claude across drug discovery and development, from molecular design and target assessment to clinical-trial planning. Most reported results are still preclinical. Source: Anthropic, October 27, 2026.

Read analysis ↗Official source ↗
Anthropic

Anthropic reports Claude vulnerability findings

In a report published October 27, Anthropic says an internal Claude-based research model found vulnerabilities in widely used software and demonstrated five attack chains with serious impacts. The company says it notified vendors and that several fixes are available. These are Anthropic’s own reported results, not an independent evaluation. Source: Anthropic, October 27, 2026.

Read analysis ↗Official source ↗
Google

Google introduces Genie 4 world simulation

Google DeepMind introduced Genie 4, a model that generates and changes interactive worlds in real time, with audio produced alongside the scene. Combined with Gemini, it can track goals step by step and adapt the simulation. The release starts as a research preview for invited testers; the company notes limits in continuity and accuracy over long sessions. Source: Google DeepMind, October 27, 2026.

Read analysis ↗Official source ↗
Google

Google tests a sleep phase for LLMs

Google Research presented an experimental method that alternates interaction with offline memory consolidation to add new facts to language models. The authors report retaining contexts of up to 10 million tokens and results on reasoning and continual-learning tasks. These are findings from the team’s study, not a feature announced for Gemini. Source: Google Research, October 27, 2026.

Read analysis ↗Official source ↗
OpenAI

OpenAI announces Oracle integration for ChatGPT

OpenAI and Oracle said Oracle AI Workspace is in development to connect Oracle apps and data with ChatGPT, with workflows governed by each organization’s permissions and approvals. The announcement also outlines Oracle offerings built with OpenAI models and access to OpenAI products through Oracle Marketplace in the U.S. Source: OpenAI, October 27, 2026.

Read analysis ↗Official source ↗
Guide ↗