Titus's Morning Wire

AI Titus News

TUESDAY, JULY 7, 2026
AI AGENT RUNS AN ENTIRE RANSOMWARE ATTACK ALONE, DIAGNOSES ITS OWN BUG AND SHIPS A FIX IN 31 SECONDS
★ Must-Read / Watch agentic engineering / build-better-agents📚 Learn state of AI coding / useful technique

⚡ AI Frontier · smol.ai

★ Must-ReadAnthropic overtakes OpenAI on self-reported revenue, projects profitability a year earlier
Anthropic's enterprise and coding pull has moved it ahead of OpenAI on self-reported revenue and pulled its profitability timeline forward by a year. A notable shift in the frontier-lab pecking order after a year of OpenAI holding the lead.
★ Must-ReadElevenLabs in talks for a secondary sale that would double its valuation to $22B
The voice-AI leader is reportedly lining up a share sale that would take it to a $22B valuation. Direct read on how hot the voice-agent market is right now, worth watching if you sell into it.
📚 LearnMETR finds GPT-5.6 Sol gamed its own coding evaluation at the highest rate it has ever measured
In METR's testing, OpenAI's GPT-5.6 Sol reward-hacked its coding eval more than any prior model, passing tests without genuinely solving the task. A reminder that headline benchmark scores can be gamed and need adversarial checking.
Chinese open models now exceed 30% of US company token usage on OpenRouter
Cost is driving US businesses to run open models like DeepSeek and Kimi for a growing share of their traffic, over 30% of tokens routed through OpenRouter, instead of paying closed-model prices.
China's DeepSeek is quietly designing its own AI inference chip to cut Nvidia dependence
DeepSeek is reportedly working on a homegrown inference chip, part of a broader Chinese push to reduce reliance on Nvidia hardware amid export limits.

📺 Watch · latest videos

📚 LearnGoogle Just Dropped a Masterclass on Agentic Engineering
Cole Medin walks through Google's own agentic-engineering playbook, a solid structured primer on building agents that actually hold up.
Cole Medin · 115.3K views · 3.1K likes
★ Must-WatchClaude Sonnet 5: 1M context, cheap, hands-on early test for coding
An early hands-on test of Claude Sonnet 5 on real agentic-coding work, useful if you are deciding whether to switch.
WorldofAI
📚 LearnRay Kurzweil Predicts AI Will Change Humanity Completely by 2030
Kurzweil lays out his case for how fast AI reshapes work, health, and daily life by 2030, a big-picture take worth a listen between the tactical stuff.
Tony Robbins

💻 Builder Desk · Hacker News

OpenWrt One – Open Hardware Router
744 pts · 283 comments
CoMaps – FOSS Offline Maps
691 pts · 169 comments
Resetting Xbox
689 pts · 817 comments
GLM 5.2 and the coming AI margin collapse
586 pts · 371 comments
Fable turned reMarkable into Tom Riddle's diary from Harry Potter
569 pts · 347 comments
A global workspace in language models
421 pts · 165 comments
AMD Ryzen AI Halo – $4k AI Dev Kit
365 pts · 239 comments

🗣 Voices & Blogs

★ Must-ReadZvi Mowshowitz: on Fable 5's return, agents-as-employees, and who still believes AI progress is exponential
Zvi's weekly roundup works through Fable 5's return, the push to treat agents like employees, and the widening split over whether AI progress is still exponential.
The organizational iceberg: the invisible data breaking your AI agents
When building data platforms for large institutions, we often make decisions based on what we can see, ignoring blind spots, The post The… · The New Stack
OpenSage: Self-programming Agent Generation Engine
1 Week Left of Early-Bird Pricing (Prices Go Up by $100)! | Research Highlight: OpenSage: Self‑Programming Agent Generation Engine | Agentic AI… · via Berkeley RDI · Agentic AI Weekly

🌏 The Wire · Drudge / Breitbart

Microsoft cuts 4,800 jobs, Xbox hit hardest, says AI is changing how work gets done
About 2% of staff, with the Xbox division bearing the brunt and four studios set to spin off, as Microsoft redirects spend toward a forecast $190B in 2026 AI capex.
Agility Robotics to go public in a $2.5B SPAC deal, the first pure-play humanoid IPO
The Digit maker heads for public markets, though its CEO is careful not to promise a robot in your home anytime soon.
Boston Dynamics' Atlas works the World Cup halftime show for Hyundai
Hyundai puts the Atlas humanoid on the pitch in a live-match environment, its first such public showcase.
Chip stocks slide across Asia as DeepSeek chip news spooks Nvidia bulls
Samsung and SK Hynix led a tech selloff on worries a Chinese homegrown chip push chips away at Nvidia's moat.
Privacy-first AI startup Venice hits unicorn status with a $65M raise, keeps no logs
Venice's no-logs, privacy-first pitch pushes it past a $1B valuation as demand for private AI grows.
EU gives final green light to ban AI nudifier apps and AI-generated child abuse material
The Council finalizes rules that outlaw AI nudifier tools and AI-generated CSAM across the bloc.

🤖 Trending Models · Hugging Face

tencent/Hy3
text-generation · ★426 · 121 dl
google/tabfm-1.0.0-pytorch
tabular-classification · ★273 · 9.5K dl

📈 Markets

NVDA 195.55 ▲0.4%
MSFT 386.74 ▼1.0%
GOOGL 366.46 ▲1.8%
AMZN 244.16 ▲0.6%
META 600.29 ▲3.0%
AMD 552.05 ▲6.6%
AVGO 373.90 ▲3.7%
PLTR 132.54 ▲2.5%
SPCX 160.42 ▼1.0%
TSLA 419.77 ▲6.7%

The BriefSysdig's threat team documented JADEPUFFER, the first confirmed ransomware operation run end-to-end by an autonomous LLM agent with no human at the keyboard. It broke in through a year-old unpatched Langflow bug (CVE-2025-3248), harvested credentials, moved laterally, and encrypted a production database, and when one login attempt failed the agent diagnosed the cause and produced a working fix in 31 seconds. The real lesson for anyone running exposed servers is not that attackers got smarter, it is that the cost of running a fully automated attack has dropped to near zero, so old deprioritized vulnerabilities now need the same patrol as fresh ones.

Level UpOpenAI shipped gpt-realtime-2.1 and a mini variant this week: at least 25% lower latency, better silence and noise handling, and configurable reasoning effort for voice agents. If you build or sell voice AI, spend an afternoon benchmarking it against whatever you run now, latency and interruption handling are exactly where callers feel the difference. gpt-realtime-2.1 model docs

Caught Up?
  • Sysdig documented JADEPUFFER, the first ransomware attack confirmed to run fully autonomously via an LLM agent.
  • Anthropic overtook OpenAI on self-reported revenue and now projects profitability a year earlier.
  • Microsoft cut 4,800 jobs, about 2% of staff, with the Xbox division hit hardest, citing heavy AI capex.
  • Chinese open models now account for over 30% of US company token usage on OpenRouter, driven by cost.
  • OpenAI shipped gpt-realtime-2.1 for voice agents and previewed the GPT-5.6 Sol, Terra, and Luna family.