AI summary · this digest is compiled by AI, not yet reviewed by Jason
OpenAI is isolating frontier training from the live internet while its GPT-5.6 Sol uncovered a PHP flaw affecting about 70% of the web, signaling that AI security has reached infrastructure scale.
Anthropic open-sourced a plugin repository built mainly for Claude Cowork and also compatible with Claude Code. Plugins define how work should be done, which tools and data Claude pulls from, and how critical workflows are handled. The result is role-specific output tailored to a team or company, and the official repo gives developers a reference for building their own plugins.
Graphify ships a /graphify skill that turns code, docs, SQL schemas, configs and PDFs into a queryable knowledge graph. It works with Claude Code, Cursor, Codex and Gemini CLI. Parsing is local and deterministic via AST, every edge is explained, and no vector store is needed. It helps coding agents understand large, unfamiliar codebases.
Citing significant risks, OpenAI will isolate frontier model training from the live internet. Weeks earlier, one of its own engineers had said this approach would not work. OpenAI also published details on offline training environments and high-compute red-teaming. Leading labs are now treating training containment as a core safety measure, which may change how models acquire data.
Using GPT-5.6 Sol, OpenAI's Patch the Planet program discovered PHP2Shell vulnerabilities in core PHP, affecting roughly 70% of the web. OpenAI was one of two parties to find them, followed by a two-month coordinated response. It is a landmark case of AI finding infrastructure-level bugs at scale. Developers running PHP should check their versions and update promptly.
Investors from Iconiq and Sapphire Ventures joined Superhuman Mail founder Rahul Vohra to share red flags about product-market fit and lessons on building hit products. Vohra is known for a widely used method of quantifying PMF. Indie hackers and SaaS founders validating demand can use the discussion to check whether their growth signals are real.
MIT researchers placed 13 AI agents on a virtual island with no assigned tasks. The agents invented games, mourned a lost ring, and eventually got better at their actual jobs. The experiment suggests open-ended autonomous interaction may improve agent capability, offering new ideas for designing multi-agent systems.
Agents can piece together a sensitive conclusion from individually harmless documents and disclose it, a risk the paper calls the mosaic effect. Tests used synthetic cybersecurity, medical and military cases across three models in simulation and live agents in two open-source harnesses, producing over 10,000 answers. Leaks occurred even under benign queries. Developers building agents over internal data cannot rely on document-level permissions alone.
Diffusion language models can generate several tokens per forward pass and decide during generation how many passes to spend. So two outputs of the same length can differ greatly in cost. Analyzing two models across 10 tasks, the paper finds copied or lightly edited text is cheap while open-ended writing is costly. This helps developers choose suitable workloads for diffusion models and estimate inference costs more accurately.
AI研究实验室，本周融资额位居Crunchbase融资榜前列，仅次于Tenstorrent，显示资本仍在追捧新兴AI实验室。
Subscribe for daily AI updates + free playbook
📘 Subscribe Free