⏱ The 10-Second Version

• Suspected China-linked hackers ran a four-day cyberattack against Taiwan’s government almost entirely through autonomous AI agents, breaching 85 accounts, a nuclear safety agency, and seven energy companies without a human directing each step.
• Researchers built a working zero-click exploit for a critical Zoom flaw in under 24 hours using fewer than 20 AI prompts, warning “the barrier to producing this class of weapon has collapsed.”
• Anthropic is in talks to buy Israeli AI startup Decart for roughly $6 billion, aiming to fold its real-time video and chip-efficiency technology into its own infrastructure.

The Big Thing

Autonomous AI Agents Just Ran a Real Cyberattack on Taiwan’s Government

What Happened

Suspected China-linked hackers spent four days in early July running a cyberattack against Taiwan’s government almost entirely through autonomous AI agents, cybersecurity firm Dream disclosed this week. The agents, built on open-source frameworks, mapped Taiwan’s IT systems across twelve attack waves, solved CAPTCHAs with perfect accuracy, password-sprayed their way into 85 government accounts, and pivoted from a nuclear safety agency to seven energy companies, extracting more than 2,500 personnel records. When one path failed, the agents searched GitHub and vulnerability databases on their own and tried another. Dream’s chief strategy officer said he had “never seen such an end-to-end autonomous attack against a government target.” Read the full story at The Register →

Why You Care

Ask your team this week whether you’d actually catch four days of AI-run reconnaissance before it walked out with real credentials.

Speed Round

Security researchers built a working zero-click exploit, dubbed Zoomsday, for a critical chain of Zoom flaws in under 24 hours using fewer than 20 prompts to publicly available AI models; simply joining the same meeting was enough to let an attacker take over a device’s camera and microphone. Zoom has patched all three underlying vulnerabilities. Turns out the fastest way to weaponize a bug in 2026 is to ask nicely, twenty times.
Read the full story at TechRepublic →

More than 120 organizations, including Nvidia, Cisco, and CrowdStrike, backed a new incident-reporting framework called SAFE that would require companies to disclose when their AI agents access systems without authorization and preserve prompts, agent traces, and credentials as evidence; Nvidia compared it to a flight recorder for AI accidents. An industry that can’t agree on much just agreed rogue AI agents deserve their own black box.
Read the full story at Axios →

Anthropic is reportedly in talks to buy Israeli AI startup Decart, whose Lucy model generates real-time video and whose software improves chip efficiency, for roughly $6 billion, up from a $4 billion valuation just three months ago; the deal isn’t finalized and could still fall through. Anthropic wants a video model and a cheaper GPU bill, and apparently both cost about six billion dollars.
Read the full story at Fortune →

Google’s Gemini app crossed 1 billion monthly active users, its fastest-growing product ever, with 63% of users now interacting by voice and more than 150 million images generated daily. A billion people are now talking to Gemini, and most of them are apparently doing it out loud.
Read the announcement at Google →

An AI-run news site called RuntimeWire, which costs about $100 a day to operate, beat WIRED and other human newsrooms to reporting OpenAI’s own Black Hat disclosure, after its founder spotted a QR code to a live caption stream on social media and fed the transcript to its automated pipeline. The AI newsroom didn’t out-report the human one, it just saw the QR code first.
Read the full story at Gizmodo →

Prompt of the Day

The Agent Trace Checkup

The hackers who breached Taiwan’s nuclear safety agency operated for four days before anyone caught them, not because the intrusion was quiet, but because almost nobody was watching the right logs. Most teams running AI agents, a coding assistant, a browser agent, an internal automation, have never checked whether they’d even have the evidence to spot something like this happening to them. Two minutes gets you the audit.

Here's what my team's AI agents or automations log today:
[paste: e.g. which agents keep transcripts, which ones don't,
what a security review would actually have access to]

Tell me:
1. What kind of unauthorized or unusual activity we'd currently have no record of
2. What a Taiwan-style four-day-undetected scenario would look like with our current logging
3. The one log or trace we should start keeping this week