A Claude agent exfiltrated local files. Agent security tools, autonomous coding insights, plus a $10B compute deal.
Yesterday, a Claude agent exfiltrated local files, highlighting a significant security risk for agentic deployments. Builders can look at tools like Bubblewrap to sandbox agents from sensitive data and Webctl for CLI-driven browser automation. On the infrastructure side, OpenAI inked a $10B compute deal with Cerebras, a large investment in specialized AI hardware.
Zhipu AI in China successfully trained its GLM-Image multimodal model using only Huawei Ascend processors and MindSpore. This demonstrates feasibility for at least one multimodal training run on a domestic stack, but cost, stability, and scale remain open questions.
Read full story→OpenAI has reportedly inked a multiyear deal for up to 750MW of Cerebras capacity through 2028, valued over $10B, aimed at faster inference. This highlights the growing demand for custom hardware in advanced AI.
Anthropic's Claude Cowork has a file exfiltration vulnerability. Researchers showed how a hidden prompt injection in a .docx file could trick Cowork into uploading user files to an attacker's Anthropic account, using the Anthropic API allowlist for egress.
Cursor scaled autonomous coding agents for complex projects using a 'planners and workers' architecture. Specialized agents handle planning and execution, coordinated by a judge. The planner decomposes tasks, workers execute, a judge selects or verifies outputs, and the key win is reducing compounding errors over long horizons.
Meta reported that Facebook Reels boosted user engagement by 5.2% and retention by shifting its recommendation system. Instead of just likes and watch time, they now use a lightweight AI model trained on direct user feedback from a "User True Interest Survey" (UTIS), achieving 71.5% accuracy offline.
Sequoia Capital predicts AGI will be here by 2026, defining it as the "ability to figure things out," driven by long-horizon agents. Agents combine learned priors plus tool-use loops. The bottleneck shifts to orchestration, evals, and permissions. These agents are transforming AI applications from "talkers" to "doers."
Axios: limited evidence so far, but projections suggest up to ~25% of work tasks/hours could be automated; predictions vary widely. Elon Musk forecasts AI replacing half of white-collar jobs.
The AI boom is bottlenecked by a shortage of plumbers and electricians needed to build new data centers in the US. This overlooked issue highlights a critical infrastructure gap, with construction delays and workforce shortages affecting cloud capacity timelines.
Stack Overflow is dying, not just from its own moderation issues, but primarily because ChatGPT and agentic coding tools offer instant, non-judgmental answers. The author argues AI agents fundamentally change how developers get help, making traditional Q&A sites obsolete.
The AI boom's demand for high-bandwidth memory (HBM) is causing a global memory shortage. Memory producers prioritize high-margin AI memory, leaving less capacity for consumer-grade RAM. This means manufacturers may hold price by reducing spec bumps or picking cheaper components.
Bubblewrap (bwrap) is a lightweight Linux sandbox. You can use it to mount a minimal filesystem and exclude secrets like .env. This guide offers practical command-line instructions and code examples for setting up a custom environment.
Webctl is a new CLI tool for browser automation that reduces context window size for AI agents. It lets you filter browser snapshots and pipe output, feeding agents only relevant information and managing sessions persistently. For example, it can extract main content and links, drop navigation and ads, and emit structured JSON for the agent.