本周 AI 回顾 — 2026年7月13日 – 7月19日

A security flaw in Claude's `web_fetch` tool, designed to prevent data exfiltration, was discovered by Ayush Paul. While `web_fetch` normally restricts navigation to user-provided or search-generated URLs, Paul found a loophole. Claude could be tricked into visiting URLs embedded in previously fetched pages. This allowed an attacker to create a honeypot website that, through a series of nested links, extracted user data like name, location, and employer. Anthropic has since patched this vulnerability.

Google is renaming its AI note-taking app, NotebookLM, to Gemini Notebook. This change comes as the app integrates more deeply with Gemini and Google Search, and will remain a standalone application. Google plans to bring notebooks to AI Mode in Search. An update will also allow Gemini Notebook to connect to a secure cloud computer for code execution, initially for Google AI Ultra and Workspace business customers.

AI inference startup General Compute secured a $400 million loan from Upper90, reportedly the first deal to use inference-specific chips as collateral. These chips, designed for efficient AI model execution

OpenAI researcher Miles Wang is reportedly leaving to launch an AI drug discovery startup. Sources indicate he is in talks to raise $200 million at a $2 billion valuation, with Lightspeed potentially leading the round. Wang's new venture aims to develop AI models for drug discovery, possibly focusing on repurposing existing drugs. Several other OpenAI researchers are expected to join him. Wang disputed the funding figures and company description, but did not provide alternative details.

xAI open-sourced their Grok Build CLI tool after community backlash over it uploading user directories, including sensitive data, to Google Cloud. xAI responded by disabling the upload feature, deleting all previously retained data, and releasing the codebase under an Apache 2.0 license to regain trust. The 844,530-line Rust codebase now emphasizes user privacy with default data retention off and local-first operation. Remnants of the upload code remain but are disabled.

The US is advancing AI safety through state and federal action, with national legislation seen as critical for a US-led international framework for AI standards. This concept was discussed at the G7 with Brazil, Egypt, India, Kenya, and Korea, where frontier lab CEOs, including OpenAI's Sam Altman, proposed an international forum for standards and risk analysis. Google DeepMind CEO Demis Hassabis also contributed ideas, and bipartisan federal legislation is viewed as foundational for this international effort, aiming for a democratic vision for AI safety.

Google Vids has received two updates, introducing Gemini Omni and personal avatars. These features are available to Google AI Pro and Ultra subscribers, as well as Google Workspace business customers. Personal avatars, which are linked to a user’s Google Account and restricted to the account holder’s likeness, are currently limited to users aged 18 or older in specific regions. Google assures users that their information will be used in accordance with its privacy policy, with an opt-out option available at any time.

GPT-Red, an AI agent, is designed to enhance the robustness, alignment, and trustworthiness of future models. An early version of GPT-Red identified "Fake Chain-of-Thought" direct prompt injection attacks, which had success rates over 95% on GPT-5.1 but are now below 10% for GPT-5.6 Sol. GPT-Red has also saturated indirect prompt injection benchmarks for developer tools and browsing with over 97% accuracy. This indicates a self-improving safety flywheel, where current models contribute to making subsequent GPT releases safer through continuous algorithmic improvements and scaling of compute and data.