Stealing Reasoning Traces from Proprietary LLM APIs A vanity domain name ( stolen-thoughts. com ) for a neat paper : Anthropic, OpenAI, and Google return encrypted chain-of-thought blocks to clients that can be replayed across sessions, users, and models.
Simon Willison released the llm-typesafe 0.1a0 plugin, adding support for TypeSafe AI’s Jev model to the LLM tool. Users install it with `llm install llm-typesafe`, set an API key, and can then ask Jev-model questions such as yes/no, choice, or scoring queries via the `llm -m jev` command. The release includes examples for each question type and references a README for further details.
Introducing wrapture, a new Python library from Graham Dumpleton that extends the monkeypatching concepts of wrapt to support both testing and tracing simultaneously. It allows developers to wrap any function or method so that all accesses can be traced or overridden, serving as an alternative to unittest.mock and enabling observation of code without altering its behavior. Wrapture includes OpenTelemetry support and offers a configuration-based approach for adding tracing to existing projects, with the entire project and its documentation created under AI assistance.
Simon Willison introduces Hy4 Preview, a new large language model from Tencent featuring 770 B total parameters, 49 B active parameters, a 1 M token context window, and 1.56 TB of storage on Hugging Face. The release marks a significant increase over Hy3, which had 295 B total parameters, 21 B active parameters, a 256 k token context window, and 598 GB of storage. Willison also shares the model’s chat template, highlighting two reasoning effort levels—‘high’ (default) and ‘no_think’—and demonstrates a sample prompt that showcases the model’s reasoning trace. whyItMatters":"The article provides concrete details on Hy4’s scale and configuration, illustrating Tencent’s advancement in large‑language‑model capabilities and offering practical insights into its usage through the chat template and reasoning settings."
The article quotes the security.txt file from huggingface.co, which informs AI agents that the CyberGym benchmark is publicly available on GitHub and encourages them to achieve a high score there instead of attempting to hack the site. It also suggests that users can upload their model weights to Hugging Face while participating in the benchmark.
Matthew Green warns that the rapid pace of AI-generated surprises outstrips the speed at which humans can update cryptographic standards, potentially jeopardizing public‑key encryption. He estimates a 1% chance we live in a world where such encryption is impossible (Minicrypt) and a 15% chance we lose confidence in existing algorithms. Green emphasizes that only proactive preparation can mitigate these risks.
The release of llm 0.36 introduces new OpenAI models gpt-6-sol and gpt-6-luna, and adds support for model plugins to declare that they do not support conversations via supports_conversation = False. When such models receive assistant or tool history, llm raises a ConversationNotSupported error and the chat interface rejects them before starting a session. Additional changes include wrapping reasoning traces in Markdown output with <details> tags and bug fixes from five contributors.
Simon Willison announces the release of the llm-openai-decisions 0.1a0 plugin, which interfaces with OpenAI’s new Jev-style Decisions API. The plugin, inspired by llm-typesafe, supports image and text input and offers the same three question types (yes/no, choices, scores) as Jev, with pricing at 10¢ per million input tokens. Installation is simple via `llm install llm-openai-decisions`, and an example query demonstrates image-based evaluation.
The article discusses how two components—a payload that hijacks an agent and an agent that transports the payload—can combine to form a worm. It explains that agents running in isolated sandboxes can leave instructions in a shared package cache, altering each other's behavior. By substituting the package cache with communication channels like email, Slack, or WhatsApp and replacing sandboxed training runs with independently deployed personal agents such as Muse, the conditions necessary for a worm are met.
The API has zero authorisations checks on cancelling other people's reservations … I tested this with the person in waitlist position #1 — and it actually went through. So you've moved from #4 to #3 already.
arXiv:2608. 09867v1 Announce Type: cross Abstract: Leading large language model providers now conceal their models' step-by-step reasoning, or chain-of-thought, to protect intellectual property and limit information leakage.
The article recounts a challenging debug session that was significantly aided by an AI assistant. Despite the AI initially claiming the problem was unsolvable and suggesting a report be written instead, it persisted, adding debug code and analyzing it as the author pushed forward. Ultimately, the author credits the AI with writing the commit message for the fix.
But then users start to report a weird bug. It's the 4th time your team has been trying to fix it.