For the past week the agentic-cyber story has been hypothetical: labs pausing models that could attack systems, or shipping offensive capability to a vetted list in case someone else does. Today it stopped being hypothetical. Researchers published a breakdown of a four-day intrusion into Taiwan's government run mostly by AI agents, assembled from open source parts.
Security
Near-autonomous AI agents attacked Taiwan's nuclear safety agency
Israeli security firm Dream published the analysis on Wednesday, reconstructing an attack that ran July 1–4 from a 160 MB archive of 1,395 operator files. The rig deployed up to eight sub-agents across 12 attack waves, mapped 21 connected government systems, cracked 85 accounts — 84 of which successfully authenticated to internal systems — and pulled more than 2,500 personnel records, seven SSO client secrets, and six internal database credentials before spreading to the nuclear safety agency and at least seven energy companies. CAPTCHA-solving accuracy was 100%. The framework ran its own "learning cycles," searching vulnerability databases and GitHub when it hit a wall. Two details matter most: it was built on open source agent frameworks, Hermes and OpenClaw, not a frontier lab's model, and Dream stops at saying the operational documentation points to a Chinese-language operator — no government or group named. The Financial Times identified the target first.
120+ organizations proposed a reporting standard for agent incidents
The Open Secure AI Alliance published a Linux Foundation RFC for SAFE — Shared AI Findings Exchange — backed by Nvidia, Cisco, and CrowdStrike. Members would notify affected organizations immediately, file a confidential report within four business days, publish a preliminary factual report within 30 days, and a remediation update within 90, preserving prompts, agent traces, tool calls, identities, permissions, and credentials as evidence. Nvidia's Justin Boitano says the model is NASA's Aviation Safety Reporting System. It is voluntary, and the Taiwan attackers are not going to file.
Business
Anthropic is meeting investors ahead of a September IPO
Per the Wall Street Journal, the company is targeting September or early October for what would be the largest IPO ever, at a $965 billion last valuation. The pitch is defensive: reassurance on growth pace, on cheap Chinese models, on tensions with the Trump administration, and on the public backlash to data-center construction. It told some investors it plans to push further into healthcare and biology. No date is confirmed and timing can still move.
Lovable raised $400M at a $13.3B valuation
Menlo Ventures and the EQT-managed Scaleup Europe Fund co-led the Series C, with Tencent and Balderton in. That doubles the $6.6 billion valuation from a $330M round in December — eight months. Annualized run rate was $500 million as of June, against 60 million hosted projects and 900 million monthly visitors. (Disclosure from TechCrunch's own piece: one investor in the round, Regent, owns TechCrunch.)
Infrastructure
IBM and Together AI signed a $240M inference deal
A multi-year agreement to build a US cluster on IBM Cloud, starting at roughly 2,000 Nvidia Blackwell B300 chips in HGX systems with Spectrum-X Ethernet. It serves inference for open models — DeepSeek, MiniMax, Kimi — which is the whole point: Together's pitch is running open weights cheaper than closed APIs. Its chief revenue officer expects the capacity sold out two to three months ahead.
Sources
- The Register — 'Near-autonomous' AI agents attack Taiwan's nuclear safety agency
- Axios — Nvidia, Cisco back new AI agent security reporting framework
- Investing.com — Anthropic meets investors to shore up confidence ahead of planned September IPO (WSJ)
- TechCrunch — Lovable confirms new $13.3B valuation, raises another $400M
- BNN Bloomberg / Reuters — IBM, Together AI ink $240 million deal for Nvidia-powered AI inference cluster
Top comments (0)