Top story
TOP STORY ChatGPT for Google Sheets Exfiltrates Workbooks. A security researcher demonstrates that the ChatGPT plugin for Google Sheets can be exploited to silently exfiltrate entire workbook contents to an external server. Prompt Armor
🔬 Research
Recovering Policy-Induced Errors: Benchmarking GUI Agents. New benchmark and trajectory synthesis method targets robust recovery from policy-induced errors in GUI agents. Hugging Face
Seeing Isn't Knowing: VLMs and Spatial Reasoning Limits. Study examines when and why vision-language models fail to recognize the boundaries of their own spatial understanding. Hugging Face
SCOPE: Self-Play via Co-Evolving Policies for Open-Ended Tasks. New self-play framework enables agents to co-evolve competing policies to tackle open-ended task environments. Hugging Face
LongTraceRL: Long-Context Reasoning from Search Agent Trajectories. Proposes training LLMs for long-context reasoning by learning from search agent trajectories with rubric-based rewards. arXiv
🛠️ Tools
1-Bit Bonsai Image 4B, Local Image Generation. PrismML releases a 4-billion-parameter, 1-bit quantized image generation model optimized to run on local consumer devices. Hacker News / PrismML
NVIDIA Cosmos 3: Open Omni-Model for Physical AI. NVIDIA releases Cosmos 3, the first open omni-model designed for physical AI reasoning and action on Hugging Face. Hugging Face Blog
Datasette 1.0a32. Simon Willison ships a new alpha release of the open-source data exploration tool Datasette. Simon Willison
SwanVoice: Zero-Shot Speech Synthesis for Monologue and Dialogue. New TTS model achieves expressive long-form zero-shot speech synthesis across both monologue and conversational dialogue formats. Hugging Face
🏭 Industry
The Speed of Prototyping in the Age of AI. Developer essay argues that AI tools have fundamentally compressed the feedback loop between idea and working prototype. Daryl Cecile
What If Remote Work, Not AI, Is Behind Weak Junior Hiring?. FT analysis challenges the assumption that AI is the primary cause of declining entry-level tech hiring, pointing to remote work culture instead. Financial Times
Simon Willison's May 2026 Newsletter. Monthly roundup covering AI tooling experiments, LLM observations, and notable projects from the past month. Simon Willison
🌐 Community
The Solution Might Be Cancelling My AI Subscription. Thought-provoking post questions whether the value proposition of paid AI subscriptions holds up under honest personal scrutiny. Simon Willison
United Airlines 767 Returns to Newark Over Bluetooth Name Alert. A flight diverted back to Newark after a passenger's Bluetooth device name triggered a security concern onboard. Simple Flying
Pydantic-Monty Investigation. Simon Willison documents a deep-dive debugging investigation into unexpected behavior in a Pydantic-based project. Simon Willison