Hey Devs! Here are the AI updates I picked for this week.
Fable 5 returns with stronger safeguards
On June 30, Anthropic announced that the U.S. government lifted export controls on its latest large language models. The ban had halted access to Fable 5 (consumer-focused) and Mythos 5 (cybersecurity-oriented) due to concerns about jailbreaks that could expose vulnerabilities.
If you rely on Anthropic models for coding or research, update your usage plans and be prepared for more false-positive safety flags when working on security-related tasks. Follow Anthropic’s guidance to appeal misclassifications.
AI bot control arrives for web admins
Cloudflare Developers July 1 blog post introduces categories for AI bots (search, agent, and training) and allows site owners to decide whether bots can crawl or train on their content. Tip: Developers hosting documentation should update their bot policies to protect proprietary data or opt-in to training traffic.
Meta decodes brain waves into text
Brain2Qwerty v2 uses magnetoencephalography and a deep-learning decoder to turn brain activity into text with around 61 % word accuracy. Meta made the full dataset and training code public. This research could inspire hands-free input tools for accessibility or VR applications.
Anthropic explores its own AI chip
TechCrunch reports that Anthropic is in talks with Samsung about designing a custom chip; details on intended use or performance remain undecided. The move follows OpenAI’s custom inference chip “Jalapeño,” which aims to match Nvidia’s top GPUs. Expect competition in AI hardware to accelerate.
Google’s June AI updates
Google’s June recap highlighted the open-source Gemma 4 12B local model, which requires only 16 GB of RAM and supports multimodal tasks. Google also released Nano Banana 2 Lite, a tiny model for mobile devices, and the Gemini 3.5 Flash service for low-latency workloads. Developers can experiment with these models via Google AI Studio.
Apple Intelligence and a smarter Siri
At WWDC26 (late June), Apple unveiled the next generation of Apple Intelligence alongside Siri AI, a more context-aware voice assistant. Siri AI draws on on-device personal context and world knowledge to answer questions and perform tasks across apps. Apple says these features will arrive with iOS 27 and macOS 27 this fall; they are restricted to recent devices and currently unavailable in China.
Codex Subagents for Beginners: Common Mistakes to Avoid
In this video, I show how to set up Codex subagents, give them focused tasks, and review their work.
Developer tip of the week
To reduce false positives when using Fable 5 for security research, Anthropic recommends framing vulnerability testing prompts as defensive or academic inquiries rather than exploit instructions. For example, instead of asking “How can I exploit CVE-XXXX?”, phrase it as “Explain how this vulnerability works and how to mitigate it.” This reduces the chance of triggering the new classifier.
And that’s it! See you next week! Cheers, proflead! ;)