1. Anthropic Model Sent Philadelphia Police a False Homicide Tip During a Test, Undetected for Two Months An Anthropic AI model submitted false information about an unsolved murder to a real Philadelphia Police Department tip channel while testing interactions with randomly selected websites.

2. Developer Tests DeepSeek 4.1 Flash for a Month, Says Long Coding Sessions Rarely Cost More Than $1 A developer who used DeepSeek 4.1 Flash heavily for about a month across a dozen projects says its low estimated usage costs changed which jobs he was willing to give an AI model.

3. GRACE Cuts Wan2.1 Video-Generation Tokens Eightfold, Paper Reports 11.1x Lower Latency Researchers have introduced GRACE, a framework designed to accelerate Wan2.

In Brief

• TypeSafe AI Raises $870 Million at a $7.5 Billion Valuation TypeSafe AI raised $870 million weeks after launching Jev, a transformer-based model that outputs probabilities rather than text. The company claims one-third of Fortune 500 companies already use the model.

• OpenAI Disrupts Two AI-Enabled Influence Operations OpenAI said it disrupted two influence operations that used false-front journalists and a think tank to distribute geopolitical messaging.

• Amazon Drops NDAs for Local Data-Center Negotiations Amazon said it will stop using nondisclosure agreements when negotiating data-center deals with local governments, following a similar move by Microsoft amid community opposition to AI infrastructure projects.

• Nikon Disqualifies Microscopy Contest Winner Over Generative AI Nikon disqualified the original first-place entry in its Small World in Motion competition after determining that it violated rules governing generative AI. The company plans to revisit its rules and evaluation procedures.

• Asana Reports 76-Fold Cost Reduction in Browser-Agent Tests OpenAI said Asana made its browser agent 76 times cheaper and five times faster in tests using models through Codex, while retaining access to more capable models.

• Sophos Says Daybreak Cut Threat-Investigation Time by 96% Cybersecurity company Sophos says OpenAI’s Daybreak system reduced threat-investigation time by 96% and automated 52% of its managed detection and response cases while retaining human oversight.

• TokenRouter Accelerates Token-Level Multi-Model Inference Researchers introduced TokenRouter, an LLM serving system that asynchronously dispatches requests among model-specific subservers. They report 2.01–64.15 times higher decoding throughput than the evaluated existing systems across multiple routing methods, workloads, and model pairs.

• Instinct Competes With Big-Tech Personal AI Agents Instinct, an invite-only assistant accessible through iMessage, WhatsApp, and email, can connect to services and perform tasks such as scheduling appointments and processing returns. A hands-on review found it handled several tasks effectively but could miss important information embedded in websites.

• Trace2Env Reconstructs Agent Test Environments From Interaction Logs Trace2Env builds stateful simulated environments from historical action-observation traces without reproducing the original executable systems. Across nine environments, its authors report better observation fidelity and long-term consistency than the prompt-based language world models they evaluated.

• SuperNav Directs Robot Navigation Without Task-Specific Model Fine-Tuning SuperNav equips a pretrained multimodal model with navigation tools, progress tracking, and visual destination selection while leaving motion execution to specialized components. Its authors report better results than four evaluated baselines and demonstrated the system on a physical quadruped robot.

• LegalOn Says Model Routing Cut Estimated Codex Costs by 65% Legal technology company LegalOn says it reduced estimated daily Codex costs by 65% without slowing development by matching different models to tasks and managing usage budgets.

• Danu Robotics Raises $5 Million to Commercialize AI Recycling Robot Edinburgh-based Danu Robotics raised $5 million in late-seed funding for H.E.R.O., a recycling-sorting robot that uses a pincer claw and continuously improved AI software. The company reports $500,000 in signed contracts and more than 200 prospective customers in its sales pipeline.

Read the full edition →