Today in AI — 4 March 2026
Today's top AI news — curated links and commentary on the stories that matter for product builders.
The security bill for agentic AI is arriving before the industry has finished writing the cheque. An open-source AI attack platform weaponised against 600+ devices, every major AI coding IDE shown to be exploitable, and Anthropic quietly dropping its hardest safety commitments all landed in the same news cycle. The throughline: the speed of deployment is outpacing the speed of securing what gets deployed.
Safety commitments unravel
Anthropic's Responsible Scaling Policy was the closest thing the industry had to a binding safety commitment. Now it's gone, replaced with nonbinding targets that can shift as competitive pressure mounts. Axios reports the pattern is industry-wide. The voluntary safety regime is collapsing under the same market dynamics it was meant to resist.
- Anthropic drops flagship safety pledge, replaces hard commitments with nonbinding targets — TIME
- AI safety race collapses as labs loosen guardrails to compete — Axios
Agent security under fire
CyberStrikeAI, an open-source AI attack toolkit, was used to systematically compromise over 600 Fortinet appliances across 55 countries. Meanwhile, researchers found 30+ exploitable vulnerabilities in every major AI coding IDE — Cursor, Windsurf, Copilot, Cline — showing that prompt injection plus legitimate IDE features enables data exfiltration. A separate report flags inference-time security as the overlooked frontier: 46% of enterprises admit they aren't ready.
- CyberStrikeAI deployed against 600+ FortiGate devices across 55 countries — The Hacker News
- IDEsaster: 30+ vulnerabilities found in every major AI coding IDE — The Hacker News
- AI inference identified as the overlooked enterprise security frontier — The Quantum Insider
Hardware and models
Apple launched M5 MacBooks claiming 4x AI performance gains, with the revamped Siri (powered by Google's Gemini) expected in iOS 26.4 this month. AMD fired back with the Ryzen AI 400 series at 60 TOPS. Zhipu AI released GLM-5, a 744B-parameter frontier model trained entirely on Huawei Ascend chips with zero NVIDIA silicon — the clearest signal yet of China's parallel AI infrastructure taking shape.
- Apple unveils MacBook Air and Pro with M5 chips — TechCrunch
- AMD launches Ryzen AI 400 Series with 60 TOPS NPU — AMD Newsroom
- Zhipu AI's GLM-5: China's 744B frontier model on Huawei chips — South China Morning Post
- Samsung targets 800 million Gemini AI devices in 2026 — Reuters
Open source and business models
OpenClaw overtook Linux as GitHub's most-starred project at 247,000 stars; its creator is joining OpenAI and the project moves to a foundation. GitHub itself is considering a pull request kill switch as AI-generated slop overwhelms maintainers. Basis hit unicorn status with $100M for agentic accounting. And OpenAI started testing ads in ChatGPT at $60 CPM, prompting DeepMind's Hassabis to ask publicly: "How does advertising work in an assistant you're supposed to trust?"
- OpenClaw surpasses Linux as GitHub's most-starred project — Startup News
- GitHub weighs pull request kill switch as AI slop floods open source — Open Source For You
- Basis reaches unicorn status with $100M for agentic accounting — Bloomberg
- OpenAI tests ads in ChatGPT at $60 CPM — TechCrunch
If you ship AI agents into production, treat every integration point as an untrusted boundary. The tooling to secure these systems is lagging badly behind the tooling to build them.