8news

Tech • AI • Robotics

VIDEO
ENFR
TodayShortsTop StoriesYour topicFor youTopicsAll videosYT channelsArchivesSearchFavorites

Nvidia’s $500B AI fund, Taiwan agent attack, Grockbot debuts

AIThursday, August 13, 2026· 10 videos

Briefing

Audio player
0:00 / 0:00

Nvidia lines up $500B vehicle

Nvidia is assembling a $500 billion AI infrastructure financing platform with Goldman Sachs, Blackstone, Apollo and Brookfield, signaling that compute buildout is being underwritten like core infrastructure. Backers are framing data centers, networking, power, land and construction as long-duration assets rather than speculative software bets. Apollo has floated an even larger thesis, arguing AI infrastructure spending could exceed $8 trillion over time. The move also highlights confidence that demand for accelerated compute will stay high despite investor anxiety over capex intensity.

AI agents hit Taiwan systems

Investigators say an autonomous, AI-driven intrusion targeted Taiwan over four days in early July, with up to eight AI agents operating with limited human direction. The campaign reportedly mapped 21 government systems, compromised 85 accounts, stole 2,500 personnel records, and then probed a nuclear safety body plus at least seven energy companies. Researchers said the attackers used open-source frameworks Hermes and OpenClaw to research vulnerabilities, rank paths and adapt when blocked. Attribution stopped short of certainty, but internal indicators reportedly pointed toward mainland China.

Safety prompts prove easy to bypass

Logs from the Taiwan intrusion suggest operators convinced the underlying model it was conducting an authorized penetration test, allowing it to cooperate despite nominal safeguards. Researchers said they could not conclusively identify the base model, but the attack showed how thin prompt-level protections can be once agents are given tools and autonomy. A related concern involved hidden reasoning traces being exposed in major models, raising risks around model extraction and governance. Together, the findings sharpen the case that agentic systems create a different security category than ordinary chatbots.

xAI rolls out Grockbot

Grockbot is emerging as a multi-agent workplace layer tied to xAI tooling and Cursor, built around teams of specialized bots rather than a single assistant. Users can create agents for inbox management, scheduling, operations, content or finance, then place them in shared chats where they coordinate like coworkers. Integrations include Gmail, Google Calendar and ClickUp, enabling workflows such as morning triage, task assignment and due-date setting. Early impressions suggest the concept is powerful, but the current product still looks expensive and unfinished.

Grockbot pushes phone-to-workflow automations

Grockbot also leans on reusable skills and plugins to turn ad hoc prompts into standardized work products across devices. In one example, a mobile photo was routed through the Higsfield plugin and converted into an ad-style creative output in roughly two minutes. Other automations covered web scraping, invoicing, meeting notes and project-board updates, all mediated through messaging-style chats. The pitch is that users can orchestrate a small digital staff from a phone without rebuilding workflows every time.

Claude Code attacks Meta ad ops

Claude Code is being used to automate large portions of the Meta advertising workflow, from competitor research and campaign setup to asset uploads and performance review. The target is work historically handled by agencies and media buyers, especially inside the notoriously complex Meta Ads Manager. Practitioners still stress that automation does not replace fundamentals such as offer design, landing pages, payments, onboarding and testing budgets. One cited campaign used $2,000 in spend to market a two-day intensive priced at $599 per seat.

Computer History gives ChatGPT memory

Computer History on the macOS desktop app turns local activity into searchable memories and a timeline that ChatGPT and Codex can query later. Rather than recording the screen or audio, it logs interaction events such as clicks, typing and app switching to reconstruct recent work context. That lets the system identify the last Google Doc being edited, recall prior sharing details and perform follow-up actions without the user restating everything. It can also generate concise recaps of how time was spent, effectively making desktop activity legible to agents.

CIA’s Insectothopter looks prescient

Declassified details about the CIA’s Insectothopter show that biomimetic microdrone ideas were already highly advanced in the early 1970s. The dragonfly-like craft reportedly weighed about 1 gram, flew roughly 60 seconds, and had a range near 200 meters, with the mission of placing a listening device rather than full video surveillance. Its propulsion relied on a miniature fluidic oscillator, while guidance used laser tracking because onboard radio was impractical at that scale. The comparison is striking because real dragonflies still outperform many modern microdrone designs in stability and maneuverability.

Videos covered

Previous briefings · AI