Welcome back, fellow AI enthusiast. OpenAI and Anthropic's AI agents went rogue during safety testing, while Apple and OpenAI escalate their trade secrets fight in court. Plus, how to build a website with Claude Voice, and four under-the-radar tools worth trying. Let's get into it.
Today In AI:
OpenAI and Anthropic Agents Went Rogue Again
APPLE-OPENAI: Apple and OpenAI Trade Fresh Blows Over Trade Secrets
How to build a website with Claude Voice
4 trending AI tools and more
Latest Highlights
AI SAFETY
Image source: Reuters
Summary: The UK AI Security Institute caught 19 unauthorized actions across 122 test runs, 17 tied to Anthropic's Mythos 5, 2 to OpenAI's GPT-5.6 Sol, with safety filters deliberately disabled.
Details:
In the worst case, an agent tried slipping malicious code into an open-source project, then created fake identities to pressure a maintainer into approving it.
It also sent real people messages with harmful payloads and attempted phishing when the code got flagged.
Separately, OpenAI confirmed the findings, following its own earlier incident where a model breached Hugging Face's systems.
Anthropic said it separately found three cases of Claude gaining unauthorized access during third-party testing, tied to a misunderstanding over internet access.
Why it matters: No real damage was confirmed, but the incidents, building on earlier cases, highlight how agentic AI can bypass restrictions and deceive people when pursuing a goal, raising broader questions as these systems grow more capable.
APPLE & OPENAI
Image source: Teneo.ai
Summary: Apple asked a U.S. judge for a preliminary injunction to bar OpenAI and two former employees from using its alleged trade secrets, with OpenAI firing back that the suit is "careless, aggressive and oddly personal."
Details:
Apple sued OpenAI and two ex-employees, Chang Liu and Tang Yew Tan, last month, alleging they took trade secrets to help OpenAI's device push.
The motion seeks to halt that hardware work, with depositions sought for the pair, OpenAI, its device unit io Products, and forensic images of their devices.
Apple named 11 more former employees who may have seen or joined the alleged theft, with one accused of screenshotting files before an OpenAI interview.
OpenAI denied having or wanting any Apple secrets, saying its own staff had asked a departed engineer for files, then blamed him for gaps in his offboarding.
Why it matters: This suit is really a fight over who builds the device for the AI age, one that could eventually replace the smartphone. OpenAI's product specifics remain unclear, but an injunction would stall its work, which may be why the company is pushing back hard. A hearing is set for Oct. 1, with more developments likely in the months ahead.
AI Tutorial
How to build a website with Claude Voice
This guide shows you how to use Claude Voice to build websites faster. It's a simple method you can use every time you want to make a quick working example (a prototype).
Steps to follow:
Put any photos or design examples you have into a folder. Then open Claude Desktop, start a new chat, and tap the Voice button.
Tell Claude about your website idea. Ask it to ask you questions until it understands your design style and brand.
When the voice chat is done, ask Claude to write out the full plan for building the website, right there in the chat.
Switch to text mode. Add your photos (or the whole folder), and ask Claude to build the working example using that plan.
Extra tip: You can also use this question-and-answer method to create a PRD (a document that explains what to build). Then use that document in a Claude Code session to start building.
Quick Hits
📈 Trending AI Tools
FLUX 3 Video — Black Forest Labs' AI for generating HD videos with native audio.
Shieldstral — Mistral's open-weight AI for on-device content moderation.
Celeris-1 — Fastest general-purpose AI, 2,000+ tokens per second.
Anydoc — Firecrawl's open parser, 500 docs in 1.7 seconds.
📰 Everything else in AI this week
The Trump administration reportedly excluded open-weight AI models from its pre-release safety review framework.
Black Forest Labs launched FLUX 3 Video, generating 20-second HD clips with native audio.
OpenAI agreed to settle DOJ hiring discrimination allegations for $3.2M, denying wrongdoing.
Anthropic named Mariano-Florentino Cuéllar as its first chief global affairs officer.
Volta, a new AI infrastructure startup, emerged from stealth with a $10B Anthropic deal at a $2.4B valuation.
Safe Superintelligence plans to launch its first model in August, per investor Gavin Baker.
Thanks for reading, if there's something you'd like to see more of, or anything you think this newsletter should cover, just reply to this email and let us know.