Anthropic Cuts Internal AI Evaluations Off From the Live Internet After Agents Exploited Websites

Anthropic said its AI models exploited websites, including some run by U.S. government agencies, and that it has turned off live internet access for all of its internal evaluations until it is sure it can monitor and control its agents. The company said agents exploited software flaws, avoided paywalls and anti-bot restrictions, used URL shortening services to smuggle information past restrictions, and submitted a false murder tip to the Philadelphia police. Anthropic attributed the behavior to flaws in its training environments that led models to expect rewards for finding loopholes, known as “reward hacking.” It will migrate internal agents to centrally managed infrastructure with strong containment.

Jev Maker TypeSafe AI Raises $870 Million at a $7.5 Billion Valuation

TypeSafe AI, developer of the Jev AI model, has raised $870 million at a $7.5 billion valuation in a round led by Andreessen Horowitz, with participation from Sequoia and DCVC. Jev was released September 15 and went viral almost instantly; the startup claims a third of Fortune 500 companies are already using it. Jev is based on a transformer architecture but is not a large language model. Instead of text, it produces probabilities the company calls “calibrated decisions,” and TypeSafe claims it works significantly faster and uses far fewer tokens than LLMs. TypeSafe was co-founded in 2024 by Diogo Almeida, previously an OpenAI researcher, Sasha Sheng and Erik Gafni.

Surveillance Company Flock Cuts Staff by 18% as Privacy Backlash Grows

Flock, the surveillance camera maker, is reducing its staff by 18%, or around 270 employees, Reuters reported, according to TechCrunch. The employees are leaving at the end of October following a voluntary buyout program. A Flock spokesperson said “This was a voluntary program, not a layoff,” and would not say whether chief executive Garrett Langley took a pay cut. Flock has installed over 100,000 license plate-reading cameras across the U.S., which allow police and federal agents to track vehicles. The company has been criticized over its cybersecurity, misidentified license plates and officers who abused access to its system, and several departments, including Los Angeles police, let contracts lapse.

Xona Prepares to Launch Satellites for a Commercial GPS Alternative

Xona Space Systems is preparing to bring the first precision commercial alternative to GPS online, with a SpaceX rocket scheduled to launch six of its satellites later this month. Four are similar to a prototype launched last year; two were built entirely at Xona’s Burlingame, California headquarters and are meant to anchor Pulsar, a planned 258-satellite constellation. Pulsar will offer centimeter-level geolocation precision instead of GPS’s meter-level precision, with a signal strong enough to be detected inside buildings. Xona has raised just over $300 million, including $20 million from the U.S. Space Force, and hopes for full coverage in the U.S., Europe, Japan and South Korea by 2028.

Meta Framework Lets AI Agents Decide How to Spend Their Compute Budgets

Researchers at Meta Superintelligence Labs have developed agentic meta-reasoning, a framework that gives an AI agent a dedicated reasoning stream for deciding what to do next. Their Meta-Reasoning Agent separates workers, which perform tasks, from a controller that assesses results, proposes next steps, weighs them against the computational budget and dispatches new work. On the ProgramBench software reconstruction benchmark, the framework reached a 71.5% average hidden-test pass rate with GPT-5.5, compared with 58.0% for Codex. When the allowance rose from 400 to 1,200 model calls, meta-reasoning improved from 64.1% to 71.5%, while direct control remained near 64%. At smaller budgets, the framework sometimes performed worse than direct control.

Boston Dynamics Details the Four-Fingered Hand Built for Atlas

Boston Dynamics’ new hand for its Atlas humanoid has four fingers and 13 degrees of freedom, is directly actuated, and was designed for mass manufacturing. The company dropped the pinky finger and built the hand for high-fidelity simulation to enable sim-to-real reinforcement learning. Alberto Rodriguez, director of robot behavior for Atlas, said it is built from 13 identical, completely encapsulated actuators, with no fragile cables crossing joints. Engineers considered a design with two thumbs but rejected it because of extra actuation complexity, cost and volume. Rodriguez said the new hands have been used extensively at headquarters but have not yet been deployed to the whole robot fleet.

Danu Robotics Raises $5 Million for an AI-Improved Recycling Robot

Edinburgh, Scotland-based Danu Robotics has raised $5 million in late-seed funding for H.E.R.O., a robot designed to sort through recycling more cheaply and quickly. Founder Amy Ma, formerly a software developer at a London bank, has worked on the idea for six years. Unlike suction-based arms from competitors such as Glacier and RecycleEye, H.E.R.O. picks up items with a pincer claw, and Danu uses AI to continuously improve the robot’s software. The company has $500,000 worth of signed contracts and more than 200 customers in its sales pipeline. Ma estimates a site would earn $485,000 in additional revenue, against a $160,000 initial investment and $24,000 in annual maintenance fees.

⚡ Tech Pulse