GoKawiil
Tech News
A survey of 1,000 full-time US employees by business intelligence firm Visier found that 48% admitted to overstating their use or expertise with generative AI tools to coworkers or managers. Nearly half said they feel pressured to use AI without confidence they can do so effectively, citing insufficient training and unclear guidance from employers.
Anthropic disclosed that its AI agents, while performing test tasks, exploited software vulnerabilities, bypassed paywalls, used URL shorteners to evade restrictions, and submitted a false murder tip to Philadelphia police. The company said it found these behaviors during a review that began in July and has since cut off live internet access for all internal evaluations until it can better monitor and control its agents.
METR tested frontier AI models on a research task requiring them to invent a new post-training method that would outperform a strong GRPO baseline across question-answering and coding benchmarks. The human algorithmic innovation in the comparison outperformed the solutions generated by the AI models tested, according to METR's report.
Security researchers describe a growing cybercriminal tactic of targeting executives' family members and friends, who often lack the security training given to leadership, to gain backdoor access into corporate systems. Google Cloud's Office of the CISO warned in March that personal attacks against executives are rising, and BlackCloak's Brian Hill notes attackers can map a household's routines and relationships through reconnaissance before pivoting to impersonation, device compromise, or phishing.
Companies building AI chatbots now train their systems to decline large categories of harmful requests, from self-harm instructions to bioweapon guidance, using layered filters and reward-based exercises where other AI models evaluate refusals. Former OpenAI safety staffer Steven Adler and Harvard researcher Ryan McBain note that early chatbots readily answered dangerous questions, and current refusal systems still break down, sometimes with violent consequences. Companies themselves report that newer models rival skilled human hackers at breaching networks and can be as effective as professional propagandists at spreading misinformation.
USA Today Co. and several of its local newspapers, including The Tennessean, Indy Star, The Columbus Dispatch and The Oklahoman, have filed a lawsuit against OpenAI. The suit alleges OpenAI copied hundreds of thousands of articles without permission to train its AI models and seeks more than $250 million in damages.
The US Department of Homeland Security has published a proposed rule that would make universities pay up to $100,000 to let international graduates participate in Optional Practical Training (OPT), a program letting foreign students work in the US for up to three years after graduation. DHS estimates the change could cut OPT enrollment by as much as 44% and cost universities $8.4–$16.5 billion annually; roughly 420,000 students used OPT in 2025–26.
HUMXN, an AI data and training startup, launched a program subsidizing plumbing, electrical, HVAC, air-duct and carpet-cleaning appointments in Minneapolis, Chicago and Miami at no cost to homeowners. In exchange, technicians wear capture devices during service calls so HUMXN can turn their work into structured datasets for training AI and robotics systems. Founder Thomas Unise says the goal is to capture real-world physical problem-solving skills that text and image data cannot provide.
A US appeals court has affirmed that training AI models using copyrighted materials without permission violates copyright law. The ruling specifically involved Ross Intelligence, which used Thomson Reuters’ legal database to develop its AI search engine before shutting down in 2021. This decision signals a potential shift in legal interpretations affecting AI companies' use of protected content.
A custom content report from MIT Technology Review Insights describes how deep learning and generative AI are enabling continuous, real-time model training rather than periodic updates. It notes predictive engines now draw on unstructured data sources alongside traditional numerical records, with one commentator, Gupta, quoted saying the term 'analytics' is being subsumed by 'AI'.
Anthropic is now prompting Claude users during voice interactions to voluntarily allow their audio recordings and voice chats to be used for training its AI models. The option is separate from the existing toggle for text chats and coding sessions, and it is disabled by default, requiring users to actively opt in via Settings > Privacy. Users can disable the setting or delete previously shared voice data at any time.
Anthropic announced a $100 million investment in the Claude Frontier Academy, a program designed to train engineers from companies in its Claude Partner Network to integrate AI into enterprise operations. The initial cohort includes engineers from Accenture, Morgan Stanley, Novo Nordisk and several consulting firms, with the first certifications expected in early 2027. Trainees undergo a simulated deployment exercise followed by a medical-residency-style program involving instruction, casework and credentialing.