Weekly Links: Growing Software, Exploits, and Disk Permissions

This week: mathematicians make a target list of problems to solve, Cloudflare starts shipping models, and Laurence Fishburn takes on the future robots.

• Do We Grow Software or Do We Design It? This is a short Tomasz Tunguz LinkedIn post, and I like the central idea. We really are getting into an era where software grows from seeds. We do need to get very good at verifying the outputs. As in nature, software systems that get too big for their architecture collapse. As in nature, ecosystems that get out of balance collapse.
• Apple changes full-disk access permissions to curb abuse from AI agents. Apple is restricting some MacOS access permission scopes in response to Meta Muse agents from the Mac app reading iMessages. Meta claims a user would have had to enable this but Apple points to the potential for abuse. Certainly users should take care what permissions they give software and especially agents on their machines. However, I really worry about this - MacOs is open in the sense that users can install things at their own risk. There are warnings in the app store etc. but they can alway be overriden. This is critical if a computer is going to be truely functional. It's a concern if this becomes a slippery slope to computers being locked down like iOS devices.
• Anthropic says GLM-5.3-Flash built a working exploit chain for $20.40. I'm scratching my head as to why Anthropic would release this information. In a nutshell they ran the slammer of the latest GLM models, demonstrate a high probably attack success rate and said they were able to produce an abliterated (no guardrail refusals) models with about $5000 of compute. Is the lesson we are supposed to take: "you don't need Mythos", "Look other models can be used to create risk, you need Mythos for defence", "Abliteration is easy, try it yourself". I think these facts were widely know, or guessed at but I' struggling to see what Anthropic wants us to learn from it.
• Gemini 4 Argon takes on GPT-6 Astra and Claude Opus 5.5 with aggressive pricing. Google hits back with it's own new model. The model version numbers are now an interesting lens on where things are for performance. Google's lead model is V4, Anthropics is V5, OpenAI's is V6...
• Trump releases AI accord with tech executives. This announcement (also in YouTube form), is another head-scratcher. The agreement seems to put in place outline agreements to carry out some of the things requested by the large AI labs in the pat few weeks. However, it seems clearly non-binding and it's wholy unclear who would need to implement this outside of the signatories.

Wishing you a great weekend!