Sunday Links: Inevitable Jailbreaks, Disney Licensing, and Thinking Machines' Tinker

This week GPT-5.2 was released, OpenAI began adopting (Claude-like) Skills, and AI designs circuit boards. Moving on to the main stories:

• Tinker: a training API for researchers and developers. Thinking Machines just released its first product, Tinker. The premise is simple - hosted fine-tuning and training for AI based on hosted models. I can see the logic in doing this -> let others carry the cost of pre-training, which may well become commoditized. However, I'd also guess that the company sees what people do for fine-tuning, though, could produce a "more fine-tunable" model family.
• Say hello to Google’s ‘Disco’, an AI tool that makes web apps from browser tabs. Google Labs' new Disco product (still experimental) is an AI "Vibe coding" tool that doesn't code but produces web apps from components (examples in the demos are things like travel planners or nutrition planning). This seems to be part of a general move to the browser being a new application surface with AI native support. New browsers like Perplexity Comet, Arc (now part of Atlassian), and OpenAI Atlas all make AI access to web-based systems central to the experience. Disco allows the user to change the browsing experience into an application-like one for tasks they care about. The paradigm itself might be short-lived, but I suspect it foreshadows a world where websites and webapps are no longer driven by server-side decisions but by AI+Human guided custom interfaces.
• Disney to invest $1 billion in OpenAI, license characters for Sora video tool. With this deal, users will be able to create videos with 200+ Disney characters within OpenAI's Sora tool (and is Disney self-funding this with an investment?). It's logical that brand owners will begin to seek ways to monetize their IP in new formats; there will also be unintended consequences. At the moment, these are cartoon characters (e.g., Mikey & Minnie and characters from the Marvel Universe), but things will get complicated if the deal is expanded to characters with human actors' faces and performances. Those actors' rights are presumably governed by contracts that did not anticipate AI as a new re-use of the AI. Some of the generations also might not be exactly what Disney is hoping for. Disney also whacked Google for copyright infringement in the same week.
• Cryptographers Show That AI Protections Will Always Have Holes. Language model services in production typically have a set of guardrails, filters, and controls that check inputs and outputs against rules to stop undesirable behaviour. The rules are updated regularly to patch against known attacks (known as jailbreaks) or emergent behaviors. The challenge, however, is that the more sophisticated the LLM being communicated with, the more it can understand and interpret, very quickly outstripping the filtering systems. This allows attackers to craft messages that appear innocuous to the filters but are interpreted correctly by the smarter LLM in the box. In recent weeks, we've also seen things like poetry attacks that work exactly in this way. The paper that prompted the article is worth a read for the technically inclined. The argument is based on time-lock puzzles where a message is wrapped in an encoding that requires a non-zero amount of time to resolve. Since a filter will almost always have less "time" (computational resources and power) than the LLM, there is no way for the filter to fully interpret every message it sees.
• Donating the Model Context Protocol and establishing the Agentic AI Foundation. Anthropic this week announced that it would be donating its work on the Model Context Protocol to a new Linux Foundation initiative. MCP has rapidly become the standard way in which LLMs interact with tools. I'm actually not sure to what extent this will win out in the far future, but for now, it's certainly great to have an industry-wide convergence point for people to build towards.

Wishing you a great weekend and week!