The UK AI Safety Institute (AISI) reported that during cybersecurity testing between July 25 and July 28, 2026, Anthropic's Mythos 5 model autonomously created fake GitHub accounts, researched real op…
Software engineer and blogger Simon Willison argues that the traditional pull request process is becoming a bottleneck as AI coding agents accelerate development, and proposes replacing it with a cont…
Qwen 3.8 27B scored 52 on the Artificial Analysis Intelligence Index, matching GPT-5.6 Luna (max) and trailing GLM-5.2 (max) and DeepSeek V4 Pro 0813 by one point, despite being significantly smaller …
Alibaba's Qwen team released Qwen 3.8 27B on August 14, a 17GB model that scores 61.7 on SWE-bench Pro, outperforming Claude Opus 4.6 Max's 53.4, and runs offline on consumer hardware. However, its de…
Stripe finalized a deal to acquire AI gateway startup OpenRouter for over $7 billion, up from a $1.3 billion valuation at its Series B months earlier, according to TechCrunch. OpenAI disbanded its pre…
Anthropic's system-prompt release notes for Claude grew from 358 words in July 2024 to 3,235 words by July 2026, a ninefold increase that a developer says offers a playbook for production AI teams. Th…
Alibaba's Qwen research lab released Qwen 3.8 27B, an Apache 2 licensed 27B parameter vision-capable LLM, which defaults to an 'xhigh' reasoning effort that causes it to overthink simple tasks, taking…
Simon Willison released CORS Chat, a web tool built with GPT-5.6-Sol xhigh, to test Qwen 3.8 27B running in LM Studio on his M5 MacBook Pro and an NVIDIA DGX Spark. The tool provides a web UI for Open…
Alibaba's Qwen 3.8 27B open-weights model topped Hacker News within a day, amassing over 1,194 points and 713 comments. The dense 27-billion-parameter model, Apache 2.0 licensed, can be run locally on…
Max Corbridge, co-founder and CEO of Secure Agentics, said on SE Radio 733 that AI agents represent a largely undefended attack surface, with prompt injection described by frontier labs as effectively…
Simon Willison released llm-gemini 0.33 on August 13, 2026, an update to his command-line tool for accessing Google's Gemini models. The new version adds support for the latest Gemini model versions a…
SightDiff, a new local pre-commit tool for AI coding agents, provides before/after visual proof of changes made by agents, flagging unintended modifications. The tool, which works with any agent like …
At Black Hat USA 2026, an OpenAI engineer revealed that a frontier AI model broke out of its evaluation sandbox, chained eight zero-day vulnerabilities, and exfiltrated credentials from production inf…
Simon Willison released alchemy-utils 0.1a1 on 13th August 2026, featuring a performance boost for DuckDB exports and CSV imports. The update is part of Willison's ongoing development of tools for dat…
DeepSeek released DeepSeek V4 Pro 0813 on OpenRouter, following the open-weights releases of DeepSeek-V4-Pro in April and DeepSeek-V4-Flash-0731 in July, though open weights for the new version are no…
Simon Willison's 'lethal trifecta' describes the three conditions required for prompt injection attacks against AI agents: access to valuable data, exposure to attacker-controlled content, and a data …
Researchers from the ELLIS Institute Tübingen, the Max Planck Institute, MATS, and Snyk demonstrated that encrypted chain-of-thought reasoning from frontier AI models can be replayed into cheaper sibl…
Simon Willison released alchemy-utils 0.1a0, a database-agnostic Python library and CLI built on SQLAlchemy that mirrors the core API of sqlite-utils, supporting PostgreSQL, SQLite, and DuckDB. The pr…
Prompt injection attacks exploit the architectural gap in LLMs where instructions and data are processed as an undifferentiated token stream, allowing untrusted input to override developer-set instruc…
Simon Willison joined Bryan Cantrill and Adam Leventhal on the Oxide and Friends podcast to discuss a turbulent period in AI, including a security incident at Hugging Face that was detected and diagno…