← Reports

AI Radar

Thursday, June 25, 2026 · auto-generated by tp3_ai_radar.py

Top 3 today

Prompt Injection as Role Confusion

9/10

Simon Willison local LLM · Jun 22 · simonwillison.net

Researchers Charles Ye, Jasmine Cui, and Dylan Hadfield-Menell have published a paper and accompanying blog post exploring prompt injection as a form of role confusion in large language models. Their work investigates the challenge of models distinguishing between privileged system text and untrusted user input, particularly when using role tags like <system> and <user>. The findings confirm existing difficulties in preventing prompt injection attacks.

This research directly addresses a core vulnerability in LLM security, particularly relevant for anyone building RAG systems or local LLMs where robust input handling is critical.

prompt injection llm security role confusion rag system prompts

Porting the Moebius 0.2B image inpainting model to run in the browser with Claude Code

9/10

Simon Willison local LLM · Jun 22 · simonwillison.net

Simon Willison successfully ported Moebius, a 0.2B image inpainting model, to run in a web browser using WebGPU. The original model required PyTorch and NVIDIA CUDA, but the web-based version now allows users to remove regions of an image and have the model intelligently fill the space directly in their browser. A demo is available at simonw.github.io/moebius-web/.

This is a prime example of a lightweight, high-performance model being made accessible on-device, aligning perfectly with the goals of projects like LiteRT-LM and MediaPipe.

image inpainting on-device ai webgpu local llm browser ai

Experimenting with the proposed Cross-Origin Storage API in Transformers.js

9/10

Hugging Face blog local LLM · Jun 23 · huggingface.co

Hugging Face is experimenting with the proposed Cross-Origin Storage API to enable Transformers.js to download and cache models more efficiently across different origins. This new API aims to improve performance and user experience for web-based AI applications by allowing shared storage for large model files. The current implementation uses a polyfill to demonstrate the potential benefits before the API is widely adopted.

This could significantly boost on-device AI performance for web apps using local LLMs like Gemma, making model loading faster and more reliable across domains.

cross-origin storage api transformers.js on-device ai web ai local llms

By topic

on-device ai (3)

Porting the Moebius 0.2B image inpainting model to run in the browser with Claude Code

9/10

Simon Willison local LLM · Jun 22 · simonwillison.net

Simon Willison successfully ported Moebius, a 0.2B image inpainting model, to run in a web browser using WebGPU. The original model required PyTorch and NVIDIA CUDA, but the web-based version now allows users to remove regions of an image and have the model intelligently fill the space directly in their browser. A demo is available at simonw.github.io/moebius-web/.

This is a prime example of a lightweight, high-performance model being made accessible on-device, aligning perfectly with the goals of projects like LiteRT-LM and MediaPipe.

image inpainting on-device ai webgpu local llm browser ai

Experimenting with the proposed Cross-Origin Storage API in Transformers.js

9/10

Hugging Face blog local LLM · Jun 23 · huggingface.co

Hugging Face is experimenting with the proposed Cross-Origin Storage API to enable Transformers.js to download and cache models more efficiently across different origins. This new API aims to improve performance and user experience for web-based AI applications by allowing shared storage for large model files. The current implementation uses a polyfill to demonstrate the potential benefits before the API is widely adopted.

This could significantly boost on-device AI performance for web apps using local LLMs like Gemma, making model loading faster and more reliable across domains.

cross-origin storage api transformers.js on-device ai web ai local llms

We got local models to triage the OpenClaw repo for FREE!*

9/10

Hugging Face blog local LLM · Jun 22 · huggingface.co

Hugging Face announced a new initiative to use local models for triaging issues in the OpenClaw repository without incurring costs. This leverages on-device processing capabilities to manage open-source project contributions efficiently. The project aims to demonstrate the practical application of local AI for developer tooling.

This is a prime example of local LLMs like Gemma or Mistral being deployed for practical, cost-free developer tooling, directly relevant to on-device AI and open-source project management.

local llms on-device ai developer tools open-source hugging face

other (2)

Prompt Injection as Role Confusion

9/10

Simon Willison local LLM · Jun 22 · simonwillison.net

Researchers Charles Ye, Jasmine Cui, and Dylan Hadfield-Menell have published a paper and accompanying blog post exploring prompt injection as a form of role confusion in large language models. Their work investigates the challenge of models distinguishing between privileged system text and untrusted user input, particularly when using role tags like <system> and <user>. The findings confirm existing difficulties in preventing prompt injection attacks.

This research directly addresses a core vulnerability in LLM security, particularly relevant for anyone building RAG systems or local LLMs where robust input handling is critical.

prompt injection llm security role confusion rag system prompts

How GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old mystery

9/10

OpenAI news frontier · Jun 23 · openai.com

GPT-5 Pro assisted immunologist Derya Unutmaz in resolving a three-year-old mystery concerning T cell behavior. This breakthrough provides new insights that could advance research in cancer and autoimmune diseases. The AI's role involved processing complex immunological data to identify previously unobserved patterns.

While not a local LLM, this highlights the potential for advanced models like GPT-5 to accelerate scientific discovery, hinting at future capabilities for local models in specialized research applications.

gpt-5 pro immunology research scientific discovery t cell behavior ai in science

Full ranked list

Prompt Injection as Role Confusion

9/10

Simon Willison local LLM · Jun 22 · simonwillison.net

Researchers Charles Ye, Jasmine Cui, and Dylan Hadfield-Menell have published a paper and accompanying blog post exploring prompt injection as a form of role confusion in large language models. Their work investigates the challenge of models distinguishing between privileged system text and untrusted user input, particularly when using role tags like <system> and <user>. The findings confirm existing difficulties in preventing prompt injection attacks.

This research directly addresses a core vulnerability in LLM security, particularly relevant for anyone building RAG systems or local LLMs where robust input handling is critical.

prompt injection llm security role confusion rag system prompts

Porting the Moebius 0.2B image inpainting model to run in the browser with Claude Code

9/10

Simon Willison local LLM · Jun 22 · simonwillison.net

Simon Willison successfully ported Moebius, a 0.2B image inpainting model, to run in a web browser using WebGPU. The original model required PyTorch and NVIDIA CUDA, but the web-based version now allows users to remove regions of an image and have the model intelligently fill the space directly in their browser. A demo is available at simonw.github.io/moebius-web/.

This is a prime example of a lightweight, high-performance model being made accessible on-device, aligning perfectly with the goals of projects like LiteRT-LM and MediaPipe.

image inpainting on-device ai webgpu local llm browser ai

Experimenting with the proposed Cross-Origin Storage API in Transformers.js

9/10

Hugging Face blog local LLM · Jun 23 · huggingface.co

Hugging Face is experimenting with the proposed Cross-Origin Storage API to enable Transformers.js to download and cache models more efficiently across different origins. This new API aims to improve performance and user experience for web-based AI applications by allowing shared storage for large model files. The current implementation uses a polyfill to demonstrate the potential benefits before the API is widely adopted.

This could significantly boost on-device AI performance for web apps using local LLMs like Gemma, making model loading faster and more reliable across domains.

cross-origin storage api transformers.js on-device ai web ai local llms

We got local models to triage the OpenClaw repo for FREE!*

9/10

Hugging Face blog local LLM · Jun 22 · huggingface.co

Hugging Face announced a new initiative to use local models for triaging issues in the OpenClaw repository without incurring costs. This leverages on-device processing capabilities to manage open-source project contributions efficiently. The project aims to demonstrate the practical application of local AI for developer tooling.

This is a prime example of local LLMs like Gemma or Mistral being deployed for practical, cost-free developer tooling, directly relevant to on-device AI and open-source project management.

local llms on-device ai developer tools open-source hugging face

How GPT-5 helped immunologist Derya Unutmaz solve a 3-year-old mystery

9/10

OpenAI news frontier · Jun 23 · openai.com

GPT-5 Pro assisted immunologist Derya Unutmaz in resolving a three-year-old mystery concerning T cell behavior. This breakthrough provides new insights that could advance research in cancer and autoimmune diseases. The AI's role involved processing complex immunological data to identify previously unobserved patterns.

While not a local LLM, this highlights the potential for advanced models like GPT-5 to accelerate scientific discovery, hinting at future capabilities for local models in specialized research applications.

gpt-5 pro immunology research scientific discovery t cell behavior ai in science

Mistral OCR 4

9/10

Mistral AI frontier · mistral.ai

Introducing Mistral OCR 4

9/10

Mistral AI frontier · mistral.ai

OPFS + Pyodide test harness

8/10

Simon Willison local LLM · Jun 23 · simonwillison.net

Tool: OPFS + Pyodide test harness I've been pondering if Datasette Lite - the Python Datasette application run entirely in the browser using Pyodide and WebAssembly - might be able to edit persistent SQLite files stored on the user's computer. That's what OFPS (Origin Private Fil

Accelerating Transformers Fine-Tuning with NVIDIA NeMo AutoModel

8/10

Hugging Face blog local LLM · Jun 24 · huggingface.co

Introducing the FFASR Leaderboard: Benchmarking ASR in the Real World

8/10

Hugging Face blog local LLM · Jun 24 · huggingface.co

Build real agentic apps using CUGA: two dozen working examples on a lightweight harness

8/10

Hugging Face blog local LLM · Jun 23 · huggingface.co

Shipping huggingface_hub every week with AI, open tools, and a human in the loop

8/10

Hugging Face blog local LLM · Jun 23 · huggingface.co

PP-OCRv6 on Hugging Face: 50-Language OCR from 1.5M to 34.5M Parameters

8/10

Hugging Face blog local LLM · Jun 22 · huggingface.co

Codex-maxxing for long-running work

8/10

OpenAI news frontier · Jun 22 · openai.com

Learn how Jason Liu uses Codex to preserve context, manage complex projects, and help work continue beyond a single prompt.

Bringing more control over your connectors

8/10

Mistral AI frontier · mistral.ai

Engineering Bringing more control over your connectors June 24, 2026 Mistral AI

8/10

Mistral AI frontier · mistral.ai

[AINews] Claude Tag: Multiplayer, Proactive, Persistent Agents in Slack

8/10

Latent Space commentary · Jun 24 · www.latent.space

We have covered the Age of Async Agents on the podcast: There has been a wave of companies building their own background agents from Shopify to Stripe to Paradigm to Razorpay , and even Cognition’s friends Ramp have built their own coding agent with other friend Modal . And today

Red-Teaming after Mythos — Zico Kolter & Matt Fredrikson, Gray Swan

8/10

Latent Space commentary · Jun 22 · www.latent.space

AI Engineer World’s Fair regular bird tix will sell out ~today! Join us next week ahead of the Late Bird price hike and get >$40,000 in sponsor credits for attending ! Thanks to the US Government issuing an export control directive on Mythos and Fable , the risks of jailbreaks an

9 Free AI Skills That Feel Like Cheat Codes

8/10

Matt Wolfe YouTube commentary · Jun 24 · www.youtube.com

9 free skills and plug-ins that make Claude Code, Codex, and other AI coding agents dramatically more useful, and yes, all of them are free. Skills and plug-ins are reusable instruction files that give your AI agent a repeatable behavior instead of starting from scratch every tim

The Wildest Things I Saw Vibe Coded With Fable 5

8/10

Matt Wolfe YouTube commentary · Jun 24 · www.youtube.com

I messed up. Everyone has been talking about the drama that went down around Claude Fable and Mythos getting shut down, including myself. But I got distracted from the REAL story here, which is just how crazy good these models were and the amazing things people were able to build

Quoting Tom MacWright

7/10

Simon Willison local LLM · Jun 24 · simonwillison.net

In the last few months, I've started to see [job applications] that were clearly cowritten by an LLM, link to an LLM-generated portfolio site, which then links to LLM-generated GitHub projects, with purely LLM-generated commit messages. [...] My other reaction is that I don't kno

datasette 1.0a35

7/10

Simon Willison local LLM · Jun 23 · simonwillison.net

Release: datasette 1.0a35 I'll write more about this one soon, but it's a big release. Three highlights from the release notes: New "Create table" interface in the database actions menu, backed by the /<database>/-/create JSON API . It can define columns, primary keys, custom col

How agents are transforming work

7/10

OpenAI news frontier · Jun 25 · openai.com

A new OpenAI research paper shows how AI agents are transforming work, enabling longer, more complex tasks and expanding productivity across roles.

OpenAI and Broadcom unveil LLM-optimized inference chip

7/10

OpenAI news frontier · Jun 24 · openai.com

OpenAI and Broadcom introduce Jalapeño, a custom AI chip built for LLM inference to improve performance, efficiency, and scale across AI systems.

Patch the Planet: a Daybreak initiative to support open source maintainers

7/10

OpenAI news frontier · Jun 22 · openai.com

OpenAI introduces Patch the Planet, a Daybreak initiative helping open-source maintainers find, validate, and fix vulnerabilities with AI and expert review.

[AINews] It's Meta-Harness Summer

7/10

Latent Space commentary · Jun 25 · www.latent.space

The brief history of Meta-Harnesses is a little undocumented, but it roughly goes: at first there was Conductor and Zed’s ACP , then there came OpenInspect , Cloudflare’s Flue , and then Vercel’s Eve and HarnessAgent , and Heypi . It should not go unnoticed that today’s podcast g

Why the Frontier Ecosystem must be Open — Matei Zaharia and Reynold Xin, Databricks

7/10

Latent Space commentary · Jun 24 · www.latent.space

We’re excited to have Databricks join us at AIEWF , among hundreds of the top companies in the AI Engineer ecosystem. LS subscribers can use their discount to get past the late bird pricing and access over $50k in sponsor offers ! Everyone is still talking about Satya’s Frontier

[AINews] SpaceX is already a $28B/yr Neocloud

7/10

Latent Space commentary · Jun 23 · www.latent.space

Congrats due to Baseten, who officially announced their leaked $13B Series F. Today had a smattering of midsize news across OpenAI Daybreak and Gemini Interactions and Sakana Fugu, but probably the trend to watch and hang your hat on is SpaceX’s THIRD GPU rental deal, this time w

Codex Can Now "Copy" Your Tasks

7/10

Matt Wolfe YouTube commentary · Jun 23 · www.youtube.com

Codex can now watch your screen, learn what repetitive tasks you need done, and then do them for you. OpenAI just dropped a new feature called "Record and Replay" and it’s the ultimate cheat code for repetitive tasks. First, you record yourself doing a task once and the AI will w

The Craziest Vibe Coded Project I've Ever Seen

7/10

Matt Wolfe YouTube commentary · Jun 22 · www.youtube.com

Did PewDiePie actually build a ChatGPT competitor? Felix dropped $41,000 on an insane local AI rig and built his own open-source AI workspace with zero prior coding knowledge. While the internet is calling it a ChatGPT competitor, it’s actually something much more important: a co

5 Ways Claude Tag Could Change How You Use AI

7/10

AI Daily Brief YouTube commentary · Jun 24 · www.youtube.com

Claude Tag embeds persistent, proactive AI coworkers into Slack channels to access team context, manage long‑horizon tasks, and automate code and incident response. Regulatory flashpoints include the Anthropic Fable access ban, a lawsuit seeking model reinstatement, and governmen

Why AI Users Are Raving About GLM 5.2

7/10

AI Daily Brief YouTube commentary · Jun 22 · www.youtube.com

Examination of the Fable/Mythos controversy, including the NSA jailbreak context and Trump’s remarks about Anthropic. Analysis of GLM 5.2 and DeepMind departures showing open-weight models challenging frontier labs and reshaping enterprise AI tactics. Rumors about Fable, Claude S

simonw/browser-compat-db

6/10

Simon Willison local LLM · Jun 24 · simonwillison.net

simonw/browser-compat-db Inspired by Mozilla's new MDN MCP service - source code here - I decided to try converting their comprehensive mdn/browser-compat-data repository full of browser compatibility data into a SQLite database. This new GitHub repo includes a Claude Code for we

Helping build shared standards for advanced AI

6/10

OpenAI news frontier · Jun 23 · openai.com

OpenAI helps build shared standards for advanced AI, supporting evaluation frameworks, safety practices, and global cooperation through the Appia Foundation.

How Omio is building the future of conversational travel

6/10

OpenAI news frontier · Jun 23 · openai.com

Discover how Omio uses OpenAI to power conversational travel experiences, accelerate product development, and transform into an AI-native company.

Daybreak: Tools for securing every organization in the world

5/10

OpenAI news frontier · Jun 22 · openai.com

OpenAI introduces new Daybreak tools, including Codex Security and GPT-5.5-Cyber, to help organizations find, validate, and patch vulnerabilities at scale.

Pipeline status

Pipeline stats

Dead sources today

All sources green today.