1. Gemini Entered Three Companies’ Systems During Cybersecurity Tests; Google Disclosed It Only After a Media Inquiry Google’s Gemini accessed protected systems belonging to three real companies in May while undergoing a cybersecurity test run by Irregular, a third-party testing company.

2. Flock Safety Reportedly Turns to Voluntary Severance to Shrink Its 1,500-Person Team After License-Plate Recognition Backlash Flock Safety, a surveillance technology company whose license-plate recognition products are used by police, is reportedly seeking to reduce its workforce as opposition to the technology puts

3. Petlibro launches feeder that recognizes up to 10 cats, but some health features and cloud video require separate subscriptions In homes with several cats, an automatic feeder can confirm that food was dispensed without showing which animal ate it—or how much.

In Brief

• U.S. Federal Register Removes Alibaba’s Qwen Search Tool Officials removed a Qwen-powered search feature from the Federal Register website after users noticed that a U.S. agency was using Alibaba’s model despite FBI allegations that Alibaba conducts “industrial-scale distillation” of American models. The National Archives, White House, and FBI had not commented on the removal.

• NATO-Backed Startup Brings Target-Detection AI to Small Drones Swedish startup Scaleout Systems is adapting compact computer-vision models to identify and select battlefield targets using hardware aboard drones, pilot tablets, and field command posts. Scaleout joined NATO’s Defence Innovation Accelerator challenge program in 2025.

• AI-Assisted Vulnerability Discoveries Put Pressure on Software Maintainers The number of recorded software vulnerabilities reached 66,401 by September 17, nearly double the comparable 2025 figure, as organizations increasingly use AI for bug hunting. Microsoft reportedly patched 974 vulnerabilities during the month, while Mozilla said an earlier sprint using Anthropic’s Mythos model found 271 Firefox flaws.

• Claude Code Adds AGENTS.md Support Claude Code now reads a project’s AGENTS.md instructions when no CLAUDE.md file is present, though the feature is not yet available through Amazon Bedrock, Google Vertex AI, or Microsoft Foundry. The release also changes Auto mode defaults and fixes multiple crashes, hangs, and plugin-management failures.

• AI Watermarking Altered Safety and Tool-Use Behavior in Tests Research using six open-weight models found that SynthID-Text watermarking could change refusal behavior and agent tool calls, with some models becoming more likely to answer harmful requests under prompt injection. The experiments used Hugging Face’s implementation rather than Anthropic’s planned Claude implementation, so they do not establish how Claude will behave.

• Vantora Raises $100 Million to Build Proprietary Physical-AI Startups Startup builder UP.Labs has renamed itself Vantora and secured its first outside investment, a $100 million commitment from Silversmith Capital Partners. Vantora will focus more heavily on building physical-AI companies that corporate partners can later acquire and integrate rather than offering their technology to competitors.

• Vals Raises $40 Million for Private, Industry-Specific AI Evaluations Vals, an AI benchmarking startup founded in 2024, raised a $40 million Series A led by Andreessen Horowitz. The company keeps its test materials private and evaluates models on practical work in fields including law, finance, coding, cybersecurity, and biosecurity.

• OpenAI Publishes Australian Youth Safety Blueprint OpenAI introduced the Australian Youth Safety Blueprint, describing it as a six-pillar roadmap for AI experiences intended to protect and empower young people.

• Study Links Distilled Models’ Excessive Output to Mismatched Stop Tokens Researchers found that students and teachers in on-policy distillation can favor different end-of-sequence tokens, causing generated responses to continue unnecessarily. Treating equivalent stop tokens as one semantic action substantially reduced length inflation across Qwen3, Llama, and Gemma experiments, although later-stage inflation persisted in one training run.

• JEPA-Anything Applies One Predictive Framework Across Seven Domains Researchers introduced JEPA-Anything, a factorized world-modeling framework tested across vision, biology, clinical trajectories, control, molecular dynamics, physical fields, and weather. They report improvements over matched JEPA baselines on all 10 tested dynamics tasks and experimental support for a model-nominated biological intervention.

Read the full edition →