Plimsoll is a self-hosted sandbox designed to run untrusted code written by AI agents while always reporting how strong the isolation around each run was. Each result carries an isolation tier (process, container, kernel, vm) and a client can demand a minimum tier; runs that cannot meet the requested floor are refused before any code executes. Runs produce cryptographic evidence: SHA-256 fingerprints of the request and the returned trajectory or result, and an external tool, plimsoll-attest, signs and replays checked runs and refuses incomplete chains. A built-in trial runner enables exam-style verification: agent-supplied programs are run against sealed scenarios (deleted files) and every run’s trajectory is hashed so identical executions match exactly. Example: a cart-pole controller (simulator compiled to WebAssembly) was run and accepted/rejected drafts produced distinct fingerprints.
Security and operability details matter: refusals that ran nothing are marked with a reason and distinct error behavior (ErrInsufficientIsolation), while errors without that mark may indicate code started and are unsafe to retry. Capability grants let sandboxed code call specified API routes without ever holding credentials, network access, or endpoint addresses; a broker outside the sandbox mints short-lived per-run credentials and enforces exact route permissions. Integrations exist for agent frameworks (Vercel AI SDK, Trigger.dev, Google ADK, Agno, CrewAI, Mastra), and a QuickJS WebAssembly provider lets users run examples locally without Docker or a daemon.
Summary generated by AI from the linked article. hn.today is not affiliated with Hacker News or Y Combinator.
Large language models (LLMs) are advanced AI systems that generate human-like text based on vast datasets. They work by predicting the next word in a sequence, enabling applications like chatbots and language translation. (kemendo.com)
Rubber Duck is an AI-powered interviewer designed to help users access and organize the knowledge stored in their minds. It aims to facilitate self-reflection and knowledge retrieval through conversational prompts. (rubberduck.talk)
![Calculate your p[DOOM], An interactive article to think critically about AI risk](https://hntoday.pixelshot.io/screenshots/20261011/2732d10502e34aca_tn.jpeg)
Calculate your pDOOM is an interactive article that encourages critical thinking about the risks associated with artificial intelligence. It explores potential societal impacts, including AI companions and their influence on personal relationships. (jesserichardson78.medium.com)
JulianFlux focuses on electrodynamic field retrieval and causal routing for AI agents, enhancing their understanding of physical environments. The project aims to improve AI decision-making by integrating electrodynamic principles into their processing capabilities. (github.com)
Jensen Huang, CEO of Nvidia, believes that AI technology is progressing well. His optimistic outlook highlights advancements in the field and Nvidia's role in AI development. (nytimes.com)
MASS employs collective multi-agent systems to optimize and evaluate AI models through recursive self-improvement, surpassing human knowledge. It uses shared intelligence as a self-supervision signal, aiming for continual learning where the model evaluates and improves itself. (pub.sakana.ai)
Today's best Hacker News stories, summarized and screenshotted, one email a day.