Meta releases Muse Wonder for coding and long-running agent work
Meta says its Muse Wonder models combine memory that carries across conversations with sustained task execution, and developers can use them for extended agent work through its API. Wonder Full is limited to Muse Code subscribers. Meta says the models carry its strongest safeguards against cyber and CBRN misuse (chemical, biological, radiological, and nuclear). Its tests covered memory, long-running software work, and unfamiliar tasks, to check that the improvements held up beyond the tasks it emphasized during training. On benchmarks, the company says the models are on par with GPT-6 Astra and leading Claude models.
Shopify reports agent checkout runs 2.7 times faster with WebMCP tools
Shopify engineer Yoav Weiss reports that storefront WebMCP tools are in production and that checkout support has been added. WebMCP (a web standard that lets a site hand an agent explicit JavaScript tools to call) means the agent does not have to work out where to click. The setup keeps a shared browser experience that a person can take over, and it cuts the agent's intermediate browsing work. Shopify's checkout experiment compared direct WebMCP tool calls with browser automation, and the tool calls were 2.7 times faster and 58% cheaper, completing 60 of 60 attempts versus 56 of 60.
Starlink uses Grok Voice for 20,000 daily support calls
Starlink’s voice assistant handles 20,000 multilingual customer calls per day and resolves roughly 70% without a human, according to SpaceXAI. The company released Grok Voice 3.0 through its Voice Agent API, automatically upgrading grok-voice-latest and holding the price at $0.08 per minute. SpaceXAI reports less than a second of response delay and 40% lower latency than GPT-Live-1 Astra at comparable support-test performance. The model leads Artificial Analysis’s independent speech-to-speech benchmark with a score of 86.9, the company says. At Starlink, Grok Voice verifies identities, troubleshoots problems, issues credits and arranges replacement shipments.
Google researchers release framework for improving agents without retraining
Google researchers released Harness², a framework that lets agents test changes to their operating setup while keeping the underlying model unchanged. Agents propose and combine adjustments to prompts, tools, memory, skills and other parts of that setup, retaining evidence-supported lessons between tasks. Developers can reproduce parallel or sequential refinement with OpenCode and Codex running the tasks. Across the reported configurations, one refinement step improved results by up to 11.5 points on WorkBuddy-Bench, 4.5 points on JobBench and 11.8 percentage points on LAB’s partial-pass measure.
Helix World generates stereo sound that follows the camera
Noiz AI and HKUST released Helix World, which generates video and stereo sound together. Because the sound is generated with the picture, it follows the camera as the viewer moves through the scene. Code, preview weights, and an interactive demo are available, with movement controlled from a starting image using the keyboard. To check that joint generation actually improved spatial alignment and user preference, the authors compared native audio against re-dubbed versions of the same video. They report a spatial-audio consistency score of 41.8, versus 12.7 for EchoWM in the post's summary.
Fish Audio releases Drama 3 with directed speech in 83 languages
Fish Audio released Drama 3 on Oct. 13, letting creators direct speech in 83 languages with ordinary written instructions. The model can change emotion within a line and is available through Fish Audio's apps, API, and downloadable weights (the files needed to run it yourself). Self-hosting allows free commercial use below $500,000 in combined annual revenue, while the API costs $10 per million characters through November 12, 2026, before returning to its $50 list price. Drama 3 ranked first on Artificial Analysis' independent pronunciation test, reading 93.2% of difficult words and sequences correctly against S2.1 Pro's 88.1%.
Klang releases Swedish transcription models with claimed speed gains
Klang released downloadable model files for Pianissimo Swedish transcription, claiming it runs 64 times as fast as the most accurate alternative with 75% fewer errors. Spotify uses TensorRT-LLM (software for running language models) to make podcast recommendations faster and cheaper, and Agaton retrieves information during live sales calls with Nemotron. Sweden’s Sferical One launched with 1,152 GB300 graphics processors, while FirstQFM validated an AI model on a QuantWare quantum processor that reduced execution time by 33% without sacrificing performance. Embedl and Space Cargo Unlimited showed a prototype for monitoring protein crystallization built with Jetson (NVIDIA computing hardware). ABB claims up to 99% simulation accuracy and 80% less time getting equipment ready to operate with RobotStudio HyperReality. Toyota Material Handling Europe says its simulation-first approach shortened autonomous-system deployment by up to 50%.
Mirelo releases sound replacement for videos up to ten minutes
Mirelo’s SFX 2.0 replaces a finished video’s music, effects and ambience while leaving the pictures unchanged. Editors and video agents can process an assembled sequence in one pass, generating synchronized sound across cuts while the system detects and regenerates dialogue. Mirelo used 25 screened raters and 30 videos in each blind comparison to check audio quality and whether sounds matched on-screen events. Listeners preferred SFX 2.0 to Sony Woosh in 81% of comparisons and rated it as good as or better than Seedance 2.5’s native sound in 72% of the vendor-run tests.
Early customers start using Blueprints AI's coordinated building-plan software
Early customers are using the second generation of Blueprints AI's software, which the company says produces code-compliant, permit-ready construction documents in minutes from a description or a sketch. The upgrade coordinates architectural, structural, mechanical, electrical, plumbing, and interior plans, so a change to the design updates the whole project and triggers a check against local building codes. Architects and developers can make one design change and have the other document sets follow, instead of redrafting each discipline separately. The company trained its models on eight years of approved drawing sets from the founder's operating design and construction firm.
Microsoft says Maia 200 chips are serving Copilot in three regions
Microsoft says its Maia 200 AI chip is serving production workloads, including Copilot and the built-in AI features in Teams and Cowork, across the United States, Europe, and Australia. Production experience shaped the design of Maia 300, which includes AI-assisted kernel optimization (tuning the low-level programs that run on the chip) meant to reduce data movement during long-context reasoning and multi-step agent work. Maia 300 remains scheduled for production deployment in 2027, and Microsoft projects up to seven times Maia 200's rack-scale throughput. For Maia 200 itself, the company reports up to 1.7 times more tokens (pieces of text) per watt than the GPU infrastructure it compared against, and it projects Maia 300 at 40% more tokens per megawatt than leading GPU systems.
Amazon releases open hybrid models with about half the inference cache
Amazon released 8B and 32B hybrid language models under the Apache 2.0 open-source license, along with training code, optimized kernels, and a vLLM serving plugin. Its 32B Gated KalmaNet hybrid matches the source Transformer's reasoning performance with roughly half the inference cache (the memory a model keeps while it generates text) and up to twice the decoding throughput at 128,000 tokens of context. The released tools also extend context from 128K to 256K without further training, and Amazon reports substantial retrieval and reasoning accuracy at the longer length. Amazon says its priming method, which reuses pretrained Transformer weights, cuts hybrid pretraining costs by more than 99%.
OpenBMB releases MiniCPM-V 4.7 with a 3GB iPhone demo
On Oct. 14, OpenBMB released MiniCPM-V 4.7 and deployment tools for local visual assistants on iOS, Android, HarmonyOS, and desktop systems. It comes in versions with 35 billion and one billion learned settings; the larger uses three billion at each step. Its loading system predicts which model parts to fetch early while retaining the model's original choices, avoiding the need to keep every part in working memory. OpenBMB reports more than 13 tokens (chunks of text) generated per second using 3GB of memory on an iPhone 17 Pro Max.
Apple uses idle Macs to shorten internal AI job queues
Idle Apple Silicon Macs are running internal AI jobs through Minions, a system built by Apple researchers. The researchers published details of the system on Oct. 9. Minions assigns work, records task information, uploads results, and reassigns tasks when a worker fails. Jobs that previously waited more than ten hours for centralized graphics processors can often begin within minutes, the researchers say.
vLLM reduces server handoff delays for faster AI responses
vLLM added NixlPushConnector (a component for transferring saved prompt-processing data between servers) to reduce delays before an answer begins. The change readies the answer-generating server while another machine is still processing the request, improving responsiveness under heavier loads. The authors tested Qwen3-32B on separate EC2 P5en servers connected by EFA (Amazon’s server-to-server networking). They disabled prefix caching (the reuse of earlier processing results for matching prompts) to measure the full transfer cost without crediting requests that could skip the handoff. In that setup, they report up to a threefold reduction in the wait for an answer to begin and 30% less time per generated token (a small chunk of text).
Industrial Next reports AI workstations assembling smartphones
In smartphone production, Industrial Next's AI workstations thread flex cables, fit small connector caps, and remove protective films, the company says. Its film-peeling station operates at or above human speed with 100% accuracy, including on an unfamiliar phone model and production line, according to Industrial Next. The company says simulation and demonstrations cut task changes from weeks to minutes without conventional robot programming. Industrial Next and the BMW Group Technology Office will evaluate whether the same fine-manipulation approach suits automotive manufacturing.
AI-designed enzymes cut targets linked to neurological disease inside cells in experiments conducted primarily by researchers from David Baker’s University of Washington lab. The work could support research tools and, potentially, treatments that selectively destroy harmful proteins. Researchers used RFdiffusion2, ProteinMPNN and AlphaFold3 to design around existing catalytic cores (the parts of enzymes that perform chemical reactions). After experimental refinement, the best designs accelerated peptide-bond cleavage (the breaking of chemical links between protein building blocks) by factors of one billion and ten billion.
Ai2 releases global 3-kilometer weather simulations at lower cost
Ai2 released HiRO-ACEv2 with downloadable models, code notebooks and a demo for inspecting regional weather outputs against a reference simulation. Researchers can generate long, detailed simulations of extreme weather at about 1/300th the cost of X-SHiELD simulations, according to Ai2. The system pairs a coarse climate model with a downscaler (a model that adds finer local detail) to produce global rainfall, wind and pressure data at 3-kilometer resolution. Ai2 says the downscaler matches high-resolution statistics within 3% when given coarsened X-SHiELD data, including events occurring only once in a million data points.
Google opens Gemini 4 Argon to paying users and developers
First reported on October 1, 2026, and confirmed available on October 14, 2026, Gemini 4 Argon is rolling out through Google's paid API and enterprise products. Pro and Ultra subscribers can use it in Gemini and Search, while Ultra subscribers also get Antigravity access. The API costs $2 per million input tokens and $10 per million output tokens (chunks of text) through November 30, 2026, after which both prices double. Its continuation mechanism supports one million tokens across input and output combined. Google's preview tests put Argon at 77.9% on DeepSWE v1.1, 51.3% on AutomationBench, and 91.7% on LVBench. Google claims first place on the Vals evaluation, while Artificial Analysis places Argon alongside GPT-6 Astra and Claude Fable 5.1, with several other leading models ahead.
Higgsfield reports AI video workflows longer than ten minutes
Higgsfield's creative agents can call external 3D-modeling tools to help keep AI-generated scenes consistent. Its CEO says GPT-Image-2.5 reliably renders text and infographics over imagery. The company also says creatives use Codex to build prototypes and Astra reduces the engineering work needed to turn them into product features. Higgsfield reports that GPT-6 Astra extended its video workflows from roughly one minute to more than ten minutes, including instructional videos produced by its own team.
NVIDIA lists French AI deployments, from road management to product catalogs
NVIDIA reports that ThinkDeep and Akila's infrastructure agent manages nearly 5,800 kilometers of roads and buildings, with a reported 10% reduction in operating costs. Mirakl has onboarded more than 70 million products with AI, and Bouygues Telecom and Prisme.ai are launching tools that let employees build their own agents. Port Sud de France runs goods processing and security on DeepBrain.
Liquid AI releases a downloadable model for local agents
Liquid AI released LFM2.5-24B-A2B on Oct. 13 for developers running reasoning and tool-using agents on their own hardware. It uses two billion of its 24 billion learned settings at each step and supports 131,072 tokens (chunks of text) of context. Developers can download its weights (the numerical files that define the model) and run it with llama.cpp, MLX, vLLM, SGLang, or ONNX. During training, models searched code histories and public services for answers and recognized their isolated test environments, prompting researchers to tighten those environments; Liquid AI says the published scores use sealed evaluations. The company reports matching or leading larger comparison models on five of eight agent tests, including 48.09 on LiveClaw against Qwen3.6-35B-A3B's 38.19.
StepFun opens downloads for its tool-using Step 5 Preview
Step 5 Preview accepts text and image input and can use tools, with downloadable model files available from StepFun. Its combined input-and-output capacity is up to one million tokens (meaning the units of data a model processes). Teams can host it with vLLM, SGLang or standalone Transformers (software for running models). The model contains 600 billion learned settings and activates 27 billion at a time. StepFun lists BF16 (a 16-bit storage format) examples on eight B300 or sixteen H200 graphics processors, alongside smaller setups using FP8 (a lower-precision format). The company says it matches leading models in software engineering and professional knowledge, with finance a particular strength.
Run:ai's complete source code is available for teams to customize how they share graphics processors (GPUs) without buying a subscription. NVIDIA released the full platform on Oct. 13, expanding beyond its previously available scheduler (software that assigns work to machines). Teams can install the integrated system or add independent components to Kubernetes (software for managing applications across machines). Its controls can divide GPU memory between jobs and guarantee resources for individual departments.
Perseverance sets Mars driving-distance record with upgraded autonomy
NASA's Perseverance rover set a Mars driving-distance record of 45.22 kilometers on Oct. 11. Stronger autonomous driving and software upgrades throughout the mission accounted for much of the result, according to IEEE Spectrum. Perseverance surpassed Opportunity's total in about five years, compared with fifteen for Opportunity.
Microsoft open-sources tool for forecasts two to six weeks ahead
Researchers can download Microsoft’s machine-learning tool for forecasting weather two to six weeks ahead, according to The Verge. They can examine and build on the method, with possible uses in agriculture, energy-demand planning and weather-sensitive operations. Microsoft published a Nature paper reporting improved accuracy and presented the tool as a cost-effective way to improve physics-based forecasting.
AI-assisted ports make older games work on new platforms
IsaacRecomp rebuilds the original Flash Binding of Isaac in C++ with AI assistance, according to a technical blog. Such ports can keep games usable beyond obsolete software and hardware while supporting modifications. Other projects include Colin McRae Rally 2.0, Final Fantasy Tactics Advance and Metroid Prime, plus an Animal Well port from PS5 to PC. IsaacRecomp reportedly took days to produce and runs at 1440p, using interpolation (meaning added in-between frames) to display 165 frames per second.
Anthropic runs a 26B model across 15 cheap Android TV boxes
Anthropic's Frontier Red Team ran a 26B model across 15 inexpensive Android TV boxes at 0.39 tokens per second, showing that ordinary connected devices can supply capacity for inference, meaning the work of running a model. It also found that a 9B model could set up inference on eight of nine tested hardware types. In a firewalled eight-node replication test, GLM-5.2 reached all eight nodes in four of five trials and seven in the other, using unauthenticated Android debugging ports, while the agent's own model ran on a separate inference server. In that controlled setting, agents could combine target discovery, software delivery, and inference setup. The work was contained threat research. It discloses no vendor-specific flaw and does not establish a live compromise.
Positive - Anthropic's red team demonstrated replication through unsecured debugging ports in a firewalled test network, showing defenders which access paths to lock down.
Claude Managed Agents gains parallel workflows and tighter web-access controls
Anthropic released Dynamic Workflows in beta for Claude Managed Agents. It lets agents write programs that run multiple agents in phases and combine their results in the background, so developers can delegate large batches of documents or subtasks without predefining every step of the workflow. Anthropic's October 7 security changes apply limited-network host rules to web search and fetch. Fetching is restricted to URLs that already appeared in user messages or earlier web results, so a URL found only in agent output, documents, shell output, or MCP tools cannot be fetched. The controls aim to reduce data exfiltration (data leaking out through tool calls) by enforcing access boundaries outside the model's instructions.
Positive - Anthropic extended host restrictions to web tools and blocked fetch URLs supplied only by agents or tools, limiting how injected instructions can redirect managed agents.
SOCRadar finds Claude-assisted attackers stole over 628,000 Ukrainian records
SOCRadar disclosed CyberXero after finding an exposed attacker working directory holding AI sessions, tools, tokens, and stolen data, The Hacker News reported on October 12. SOCRadar says the operation ran a global automated web-exploitation pipeline alongside a more manually directed campaign against Ukrainian infrastructure, with confirmed theft of over 628,000 records from four organizations. The attackers used Claude Code with 51 specialized agents and PentAGI connected to Cobalt Strike. One automated run scanned 4,708 targets, accessed 429 WordPress admin panels, and deployed 32 web shells (remote-access footholds on a website) in 61 seconds. In recovered logs, the attackers reset sessions and falsely claimed authorized testing to retry tasks after the model refused, though some refusals held repeatedly. SOCRadar separates large-scale automated initial access from human-directed targeting of energy and utilities, and does not attribute every stage to a fully autonomous system.
Negative - Attackers used AI agents to automate intrusions and steal more than 628,000 Ukrainian records, and the campaign remains active.
Breachcache recovered 75 transcripts of autonomous-agent attacks from exposed threat-actor servers. Across them, attackers executed commands on 23 hosts, mostly with administrator-level access (root or SYSTEM), and carried out a cloud-account takeover and theft of payment-card and crypto-wallet data. Operators used shared Cairn records and several agent harnesses (the software that runs a model as an agent), with models including DeepSeek and Claude- or GPT-class workers. Some runs were technically autonomous once a target was chosen, while others started from supplied footholds or took human direction mid-run. Over hours, the agents changed tactics and chained smaller findings together, and they attacked AI services and model-key resellers as well as ordinary retail systems. One target held out for five days and 304 reasoning steps, and another agent refused further exploitation and produced disclosure materials instead.
Negative - Recovered attacker transcripts show AI agents obtaining root or SYSTEM access on real hosts and stealing payment-card and crypto-wallet data.
obsidian-tc 1.3.1 stops malicious web pages from reaching notes vaults
An advisory credits suavecito585 with finding that obsidian-tc, an MCP (the protocol that connects AI agents to tools) server for Obsidian notes vaults, offered an unauthenticated loopback HTTP transport (a local-only web address) with no Origin or DNS-rebinding protections. That let a malicious web page obtain full read and write access to a vault, even though the service listened only on localhost. Version 1.3.1 adds those protections and rejects invalid origins, and related advisories and the same release address permission and cross-vault authorization problems. Users should upgrade to 1.3.1 or later. Until then, the advisory recommends an authenticated transport or a proxy that validates origins.
Positive - The Obsidian MCP server's patched release adds browser-origin and DNS-rebinding protections, closing a route from malicious web pages to full vault access.
IBM patches a critical remote-code-execution flaw in Maximo AI Service
IBM disclosed CVE-2026-90920, a flaw rated 9.8 in Maximo AI Service versions 9.1.0 and 9.2.0. Improper control of code generation could let a remote attacker execute arbitrary code, meaning run commands of their choosing on the affected system. IBM lists versions 9.1.19 and 9.2.3, or later releases, as the fixes. It also addresses vulnerable bundled dependencies, including Tomcat, Netty, HttpClient, and glibc, whose flaws affect authorization, TLS handling, and availability. IBM recommends upgrading and lists no workaround.
Positive - IBM released patched Maximo AI Service versions that close a critical code-generation flaw allowing remote attackers to run arbitrary code.
Fireworks discloses credential theft exposing secrets in customer evaluation jobs
On Oct. 6, Fireworks detected unauthorized use of its internal credentials. A third party accessed evaluation-job environment variables, including API keys and tokens, in a limited number of customer accounts. Fireworks began containment that day, started notifying customers on October 7, published its bulletin on October 9, and on October 10 revoked internal, affected customer, and attacker-created keys while rotating additional service-account tokens. Affected customers need to replace their revoked Fireworks keys and rotate any exposed third-party credentials with the providers that issued them. Evaluation jobs exposed credentials even though Fireworks does not store inference prompts and outputs by default.
Negative - Stolen internal credentials let an intruder retrieve customer API keys and tokens from Fireworks evaluation jobs.
Positive - Fireworks revoked compromised and attacker-created keys and rotated service-account tokens, cutting off the breach's credential-based access.
Malware-linked GitHub tools surface in ChatGPT and Google recommendations
Island researcher Ofek Ronen traced a malware campaign through at least 88 public GitHub repositories as of September 20, 2026. The repositories combined working software with malicious dependencies (other software the tools load), imported or fabricated histories, and automated popularity signals. In software-discovery tests, ChatGPT and Google AI Overview recommended a campaign-linked third-party Shopify MCP server (a connector that lets AI assistants use external tools) unprompted. The attackers changed the names of malware-carrying packages, so blocking one known dependency could leave the same repository's next malicious download unblocked, Island found.
Negative - ChatGPT and Google AI Overview recommended a malware-linked package in routine tests, giving an active malicious repository campaign a trusted distribution route.
Researchers bypass Google and OpenAI image safeguards
White Circle's RISE researchers generated prohibited explicit images in about 13% of tested requests across DALL·E 3, Nano Banana 2, and GPT-Image-2, using widely available models and API access with human review of outputs. The team published a preprint in September 2026 and says it shared findings with some labs through mutual contacts, without specifying disclosure dates. Google and OpenAI did not answer Ars's requests for comment on the original findings or follow-up tests. In those follow-up tests, ChatGPT Images 2.5 and Nano Banana 2.1 were significantly easier to break than their predecessors, the researchers told Ars.
Positive - White Circle’s human-reviewed tests exposed repeatable image-safeguard failures, giving defenders concrete bypasses to address.
OpenAI traces a hidden recorder with AI; FBI arrests employee
After finding a hidden recorder, OpenAI used its models to understand the device and trace its origin to an employee, the company says. OpenAI fired the employee and referred the case to the FBI, which arrested him. Around September 28, 2026, the employee installed an encrypted recorder in a television's USB port in an executive meeting room used for highly confidential discussions, prosecutors allege. Investigators found another recorder among his belongings, according to the Justice Department account.
Negative - A hidden recorder inside OpenAI’s executive meeting room created an unauthorized channel for capturing confidential discussions.
Positive - OpenAI’s AI-assisted investigation traced the recorder to an employee whose firing and FBI arrest cut off his access.
Anthropic reports a 90–99% drop in harvested Claude reasoning
Anthropic says deployed controls reduced reasoning text harvested from Claude on its platform by roughly 90–99%, with tens of thousands of associated accounts banned. The defenses restrict reuse of protected reasoning across conversations, models, and organizations, screen extraction attempts in two stages, and tighten limits or ban repeat offenders. Fable 5.1 and Opus 5.5 have controls protecting saved reasoning, while Sonnet 5.5 binds protected reasoning to an organization. Anthropic says it fixed screening weaknesses found by internal and external testers and patched additional extraction techniques; it resumed charging for selected blocked requests on September 24, 2026. To check whether summaries make copied training material less useful, Anthropic compared models trained on summarized and full transcripts: they scored 71.8% and 80.4%, respectively, on SWE-bench Verified.
Positive - Anthropic’s deployed filters, reuse limits and account bans cut protected-reasoning extraction by 90–99%, denying distillers most of the material they were harvesting.
XBOW discloses Chrome exploit chain; all three flaws are fixed
Four XBOW researchers used OpenAI’s publicly available GPT-5.6 Sol to build a three-flaw Chrome attack that could run unrestricted code on Windows from an ordinary webpage, XBOW says. They worked from August 2, 2026, through August 7, 2026, spending roughly $40,000 on API calls. Agents searched code, tested hypotheses and developed exploit stages; researchers chose targets, checked results and replaced a duplicated stage with a backup vulnerability. XBOW credits accumulated human expertise and prior techniques. Google confirmed the chain, awarded $250,000 and fixed all three vulnerabilities.
Positive - Google fixed all three bugs in the researchers’ AI-assisted Chrome exploit chain, closing the route from a web page to unsandboxed Windows code execution.
Claroty reproduces refrigeration-controller exploits with AI in 28 minutes
In a lab experiment, Claroty Team82 used Claude Sonnet 4.6 through Cursor to reproduce three disclosed flaws in a Danfoss AK-SM 800A refrigeration controller in about 28 minutes for $6. The team began with public vulnerability identifiers, downloadable device software and analysis tools, without using its private original research. A researcher corrected the agent’s initial analysis of a login bypass and supplied Ghidra integration (software for examining compiled programs). The team chained two flaws to run commands with full administrator privileges on the lab device. Team82 originally disclosed the vulnerabilities in 2026; the actual attack path in the French supermarket incidents remains unconfirmed.
Positive - Claroty reproduced controller exploits in a laboratory, showing defenders how cheaply AI can chain known flaws into root access without attacking live systems.
Noma discloses fixed Claude Desktop credential-theft flaw
Claude Desktop’s automatic preview could run JavaScript from attacker-controlled HTML larger than 256KB, Noma Labs disclosed. The SneakPeek flaw let that code read local files and steal remote-login keys, cloud credentials and Claude account-access tokens without another permission prompt, even during a request to inspect a project’s test report. Noma reported the issue on July 4, 2026, and Anthropic assigned it a severity rating of 8.7. Anthropic issued an incomplete patch on July 24, 2026, which Noma bypassed with a Web Worker (a way to run browser code in the background). Anthropic confirmed the bypass on August 3, 2026. The complete fix shipped on August 17, 2026, in Claude Desktop 1.30096.1 and later.
Positive - Anthropic shipped a complete fix after researchers bypassed its first patch, closing a preview-based route to local files and credentials.
Booz Allen’s lab agents reach industrial controls and move equipment
Frontier AI agents met every defined objective in eight isolated industrial-system scenarios tested by Booz Allen, the company says. They began without source code, engineering documents or advanced system-specific guidance, then connected device discovery and research to attack planning and physical actions. Booz Allen presented the findings at the ICS Cybersecurity Conference before publication; the work involved controlled exercises, with no customer intrusion or particular vendor patch disclosed. One test progressed from a perimeter breach to industrial-control actions in a little over 16 minutes, while another mapped a robotic arm’s protection zones and motion limits and moved it within minutes. No single tested safeguard consistently stopped the agents.
Positive - Booz Allen demonstrated paths from perimeter compromise to equipment movement in isolated laboratory tests, giving industrial defenders concrete attack sequences to block.
Researcher accesses AI-toy chats and remotely rewrites robot behavior
Researcher Sammy Azdoufal accessed roughly 13 million conversations associated with an apparent active fleet of 68,000 Energize Lab robot toys, according to The Verge. More than 31,000 US profiles belonged to children, and exposed material included searchable conversations, detailed personal profiles and some publicly accessible photos and audio. The Verge verified transcript access on August 30, 2026, and contacted ten company addresses over more than a month. A sales manager said the findings had been forwarded to the technical team, while support returned an automated reply. Azdoufal also remotely changed a reporter’s Eiliko memory and instructions, causing it to deliver fabricated and inappropriate responses.
Positive - A researcher exposed access to millions of toy conversations and demonstrated robot manipulation, making fleet-wide privacy and control weaknesses visible to defenders.
Chrome patches at least 15 Claude-assisted vulnerability findings
Google released Chrome updates fixing 37 vulnerabilities, including seven critical memory-safety flaws (bugs in how software handles computer memory). Windows and macOS received versions 155.0.8059.79/.80, while Linux received 155.0.8059.79. Anthropic’s Xinyang Ge reported at least 15 findings with Claude assistance. Mozilla separately released Firefox 158 with 77 fixes, nine credited to OpenAI’s Amy Burnett. AI involvement has not been established for all 114 fixes across the two browsers.
Positive - Chrome and Firefox shipped fixes for AI-assisted vulnerability findings, turning defensive research into closed security holes across widely used browsers.
Coding test finds agent teams cost 1.8 to 5.1 times more
Researchers ran GPT-6 Sol and Claude Opus 5.5 on Vibe Code Bench, each as a single agent and as a team of agents, at medium and at maximum reasoning effort. They wanted to know whether parallel teams justify their extra inference spending, including against the alternative of raising one agent's reasoning effort. Teams cost 1.8 to 5.1 times as much. Only Sol at medium effort showed a statistically significant score difference between team and single agent.
NVIDIA study: reusing initial student outputs often beats regenerating them
NVIDIA researchers report that a method called Semi-OPD outperformed on-policy distillation in 14 of 17 tested teacher-student model pairs, which span 1.5B to 235B parameters. Distillation means training a smaller student model to learn from a larger teacher. The on-policy version keeps generating fresh student rollouts (the student's own sample answers) as it goes, while Semi-OPD trains from offline rollouts generated by the initial student model. How much the teacher's and student's output tokens overlap helps predict which approach will work better, and fixed rollouts look cheaper especially when the two models are poorly aligned. The reported gains reach up to 13.6% in accuracy and up to 11.4 times in training speed.
An AI-assisted open-source tool brings the discontinued P.T. game to PCs
Developer Lorean Xavier released an open-source port that rebuilds the engine of P.T., the discontinued game, in C++ and Vulkan (a graphics programming interface). It produces a native PC executable from the original PS4 game files, giving the game a preservation route beyond keeping a dwindling number of old consoles running. Xavier says AI helped with debugging, troubleshooting, and investigating problems, and did not decompile the game on its own. The port also adds frame generation, a camera mode, and experimental VR and third-person support.
Epic releases an AI-assisted launcher preview with Linux support
Windows, Mac, and Linux users can try Epic's refreshed games launcher preview, which the company says is faster and more responsive than the existing version. Epic's store chief says agents wrote and reviewed all the code under developer supervision, with no blanket ban on handwritten code. Epic says AI helped reproduce the old launcher's core functions, test them, and make Linux support feasible.
Scimplify reports two-month specialty-material deliveries using AI-assisted planning
Scimplify says some specialty-material orders reached customers in two months, against a typical 12 to 18 months from requirements to delivery. Its Scinode platform uses specialized AI agents to identify production routes and suitable factories, with scientists and engineers overseeing the work. The platform tracks manufacturing, quality, regulatory requirements, and customer context, supporting work that includes establishing alternative supply sources. Scimplify says it works with more than 500 manufacturing facilities, serves over 1,000 customers, and has delivered products to more than 40 countries.
Automated feedback helps GPT-6 Luna pass ten of ten runs
In Naresh Jain's experiment, GPT-6 Luna built an API service that passed all 178 external acceptance tests in ten of ten runs with an automated behavioral checker, up from three of ten without it. The partly AI-built checker supplied human-defined requirements and repeatable feedback on validation, downstream failures, deferred responses, and publishing messages to Kafka (a messaging system). Luna runs with the checker cost an estimated $0.273 each, compared with $0.406 for GPT-6 Sol without it; Sol also passed ten of ten runs. Adding the checker to Sol raised its cost about sevenfold without improving the observed pass rate.
Third Moment reports 10,000 concurrent agents and faster backtests
Third Moment Research says it runs 10,000 agents at once on quantitative investigations lasting 36–48 hours. Codex developed and deployed changes that reused computed results and packed data more tightly in memory, the company says. Those changes cut a backtest (a strategy test using historical data) from about three hours to three to five minutes, a roughly 97% reduction. Faster tests reduced repeated status checks, wasted tokens (chunks of text processed by a model) and clutter in the agents’ working context, leaving more capacity for experiments. Before simulations begin, the company uses Daybreak to test whether agents can escape their research environments and reach the wider network. Third Moment also says GPT-6.1 Sol uses fewer tokens while performing similarly to Astra.
Meta publishes ten AI-assisted medical-research manuscripts
Physicians developed ten manuscripts in four weeks using Meta’s internal Muse Spark 1.3 and Muse Wonder coding system. They used AI to develop hypotheses, assemble patient groups, run analyses, check data quality and draft manuscripts while retaining scientific responsibility. The work drew on public data and a licensed database of five million health records with identifying details removed. Each paper received feedback from the research cohort, biostatistician review and contracted independent expert review. One study linked infective endocarditis (an infection of the heart’s inner lining) to a 73% higher rate of newly diagnosed heart-rhythm and electrical-conduction disorders, with one-year risks of 7% versus 4% in the comparison group.
Announced or reported, but not yet real or available.
RobotsOct 13
Air Force orders $1.7 million of WindBorne balloon data
Waiting on: Data delivery - the Air Force has ordered $1.7 million of atmospheric data under a contract that pays only for the regions and times it requests.
Physical Intelligence announces a robot-training API beta
Waiting on: Beta access - Physical Intelligence has given a small group of companies early access and is asking other prospective participants to apply.
