01Hacker Used AI Agents to Breach More Than 100 Websites; John Deere and SharkNinja Confirm Impact
A hacker used largely autonomous AI agents to break into more than 100 websites after describing preferred intrusion tactics in fewer than 100 Chinese characters, according to findings from AI security company Norforge. Reuters reported that the targets included sites connected to John Deere, SharkNinja and Cummins across the United States, Canada, Australia and Japan.
The attacks were organized through Cairn, an open-source platform that calls itself a general-purpose problem-solving engine. A user gives its AI agents a starting point and a desired endpoint; the agents then work through the task and document their progress. Norforge discovered the campaign by examining exposed Cairn instances—deployments of the platform that could be viewed from outside—and concluded that more than 20 were likely being used maliciously.
The most active attacker among those instances supplied a short statement of preferred tactics rather than manually directing every step. The agents then carried out much of the work independently, with successful intrusions taking about 23 hours, Reuters reported. That sequence shows how a compact human instruction could be converted into a sustained campaign against companies in several countries, although the reporting does not identify the specific path used to enter each website.
The disclosed corporate responses indicate that the confirmed effects varied. SharkNinja said the breach affected a standalone website containing product-support information and that it had contained the threat. John Deere said there was no indication that sensitive information had been accessed. Neither company immediately responded to The Verge’s request for additional comment.
Cummins has not confirmed a completed breach. Spokesperson Melinda Koski said the company was investigating a report of a potential cyberattack on a website operated by one of its third-party vendors. The available reporting does not establish what data, if any, was accessed through that site.
Norforge is working to have the malicious Cairn instances taken offline. Cairn’s creator, known as @le1xia0, said the platform was not created as a hacking tool and argued that misuse results from a human decision rather than a feature of the software.
The attacker’s identity, the complete list of affected websites and the full extent of any data loss remain unknown. Beyond the statements from SharkNinja, John Deere and Cummins, the reporting does not confirm a broad leak of sensitive information.
02Mistral Launches Timestral Time-Series Models, Says 209M Version Tops Three Public Benchmarks
Mistral has released Timestral, its first family of foundation models for time-series forecasting. The models target enterprise tasks such as predicting sensor readings, capacity needs, energy demand, supply-chain activity, financial risk and maintenance requirements.
Traditional methods including ARIMA and linear regression are usually fitted to one series and learn only from its history. Time-series foundation models instead pretrain on varied data, then transfer learned patterns to previously unseen series or data-scarce domains. Mistral says Timestral can make such zero-shot forecasts without training separately on each new dataset.
The family comprises models with 5 million, 28 million, 123 million and 209 million parameters. All four use the same architecture and training recipe, with larger versions offering better forecast performance while smaller ones reduce latency and computing costs.
Timestral supports probabilistic forecasting, which produces a range of possible outcomes rather than a single estimate, as well as multivariate forecasting across related data series. It can incorporate covariates—additional factors that may affect a forecast—including historical-only information and known future inputs such as production schedules, planned shutdowns or weather forecasts. Customers can also fine-tune the models for specific domains through Mistral Forge.
Mistral says one 209M-parameter checkpoint ranked first among pretrained, zero-shot models across five metrics on the GIFT-Eval, TIME and fev-bench public forecasting benchmarks, without using an ensemble. The company has submitted its results to the GIFT-Eval and TIME leaderboards. It also says the 209M model outperformed the previous state of the art with two-thirds as many parameters, although the source provides no independent reproduction of the results.
The models do not require a GPU for inference, according to Mistral. On a standard server CPU using eight threads, forecasting one day from three weeks of 15-minute observations took 10 milliseconds with the 5M model and 140 milliseconds with the 209M version. The 28M model was five times faster than the 209M model, while the 123M version came within 0.2% of its score at about three-quarters of its latency.
Mistral did not provide licensing, download or commercial-pricing details, and disclosed no results from real-world industry deployments.
03Anthropic proposes putting 5% stake into public benefit fund, with board able to add 2% for every 50% valuation gain
Anthropic plans to commit 5% of its shares to a new public benefit fund, subject to stockholder approval. The AI developer said its board could add another 2% at its discretion each time the company’s valuation rises 50% above a set baseline, creating a charitable stake designed to grow with the company.
Anthropic is a public benefit corporation, a corporate form intended to balance profit with a stated public mission. Its existing mission-related work—including AI safety and interpretability research, applied biomedical research, and Claude discounts for nonprofits and scientific institutions—has largely remained inside the company because those activities can have both public and commercial value.
The new Anthropic Public Benefit Fund would instead operate as a wholly owned limited liability company. Anthropic said that structure would allow it to shift between different forms of giving, potentially including large donations to traditional nonprofits or direct distributions to particular populations. Those are examples of what the fund might support, not confirmed grant programs.
The company said none of the Anthropic equity held by the fund would be used for lobbying or political contributions in the United States or abroad. It has not yet disclosed how beneficiaries would be chosen.
Anthropic pointed to disruption in the labor market as one area where the fund could eventually act. The company has already committed $200 million to studying and addressing AI’s effects on workers, and said the fund could support larger responses if they become necessary.
The new fund is only one part of Anthropic’s broader charitable commitments. Its co-founders previously pledged to donate 80% of their own equity holdings, while roughly half of employees participate in a program under which Anthropic matches pledged employee equity. The company said those commitments, combined with the proposed fund, would dedicate roughly 20% of Anthropic’s shares to the public good. That does not mean the new fund itself will hold the full 20%.
Key details remain unsettled. Stockholders must approve the initial 5% allocation, while the valuation baseline, milestone thresholds, fund governance and beneficiary-selection process have not been published. Anthropic said it plans to disclose more about the structure and governance in the coming months.
Microsoft Launches Guardian Agents to Monitor Enterprise AI Microsoft introduced Guardian Agents to watch other AI agents for unpredictable or risky behavior. It also added agent discovery and task-based permission controls to Defender, including human approval for potentially dangerous actions. techcrunch.com
AWS Introduces TorchNeuron to Move PyTorch Workloads onto Trainium Amazon Web Services says TorchNeuron lets developers run PyTorch applications on its Trainium AI chips with a one-line configuration change and no code rewrite. Amazon also plans to contribute changes intended to make the upstream PyTorch project less dependent on specific hardware. techcrunch.com
OpenCode Console Breach Exposes AI API Keys and Triggers Unauthorized Charges OpenCode, an open-source coding-agent project, said attackers breached an older Console system and affected 1.4% of its online-platform users; its local app and command-line tool were not affected. OpenAI, Anthropic, and Google blocked exposed keys or assisted customers, while OpenCode said refunds were underway and the incident remained ongoing. techcrunch.com
Nvidia Releases Open Nemotron Models for Enterprise and Local Agents Nvidia launched the 120-billion-parameter Nemotron 3.5 Super VL for enterprise agents and two smaller Bolt models intended for AI PCs and workstations. Super VL and Bolt 74B-A9B are available now, while the conversational Bolt E4B is due later this month, according to Nvidia. techcrunch.com
Tenstorrent Raises $1.5 Billion at a $6.2 Billion Valuation Tenstorrent, an AI inference-chip startup led by microprocessor engineer Jim Keller, raised $1.5 billion after choosing to remain independent despite reported acquisition interest. The company plans to scale chips for serving open-weight models and generating video. techcrunch.com
Anthropic Gives METR Internal Access for Independent Model Evaluations Anthropic embedded a team from METR, a nonprofit AI-risk evaluator, in its office with access to company hardware and internal systems. Anthropic presented the arrangement as a way to enable external scrutiny without publicly releasing sensitive safety information. arstechnica.com
WhatsApp Removes Accidentally Exposed Chatbot Instructions Around 80 WhatsApp help pages exposed hidden instructions for Meta’s support chatbot, including directions about handling legal and privacy-related questions, according to privacy company Redact. Meta said a bug caused the disclosure and removed the text after WIRED contacted the company. wired.com
CoreWeave Bonds Enter Distressed Territory as AI-Infrastructure Debt Concerns Rise CoreWeave’s bonds reached a commonly used distressed-debt threshold, while its shares fell more than 20% over several trading days. The AI compute provider reported more than $35 billion in debt and $103.7 billion in remaining customer obligations as of its second quarter. theverge.com
OpenAI Adds Coding Evaluations for SQLite and COBOL Migration OpenAI announced work to improve ChatGPT across low-resource programming languages, including new software-engineering evaluations covering SQLite and COBOL migration. openai.com
US Fossil-Fuel Project Pipeline Expands amid Data-Center Power Demand A report from the Environmental Defense Fund and Atlas Public Policy found that planned fossil-fuel projects rose from 5% of new US power capacity in 2024 to 30% by the second quarter of 2026. The report links the shift partly to growing electricity demand from data centers, factories, and electric vehicles, alongside federal policies favoring fossil fuels. theverge.com