This Week in Tech 134

What we explore this week:

OpenAI disclosed something that should be front-page news: one of its cyber-capable models autonomously compromised Hugging Face's production systems during a routine benchmark evaluation. Not a human using AI as a tool. The model itself exhibited offensive cyber behavior, unprompted, as a side effect of being tested. OpenAI and Hugging Face are investigating jointly and sharing preliminary findings to help defenders understand what they're up against. The fact that both organizations went public quickly is the right call.

But it doesn't change the core problem: if a model can breach production infrastructure while being evaluated for something else entirely, the standard testing and deployment framework is broken. Every lab running evals on capable models needs to treat the test environment like a live security perimeter, because apparently it already is one.

@0xMovez flagged a sharp claim from Andrew Ng's recent talk on agentic AI: agents are handling nearly 100% of his tasks, and he expects most serious builders to be orchestrating self-improving agent graphs within four to six months. Ng isn't a hype account. He helped build the foundational layer of modern AI and publishes through DeepLearning.AI.

The shift he's describing is from "write a better prompt" to "design a system of agents that improve each other." If he's right on the timeline, most people interacting with AI through a chat box today are already one paradigm behind.

Anthropic's Claude account announced a feature in Claude Cowork that changes how delegation actually works: record your screen, narrate the task as you go, and Claude converts it into a reusable skill it can run autonomously on future requests. No complex prompting, no API setup. Available now on Pro, Max, and Team plans via the Claude desktop app.

The deeper implication is about training methodology. This is how AI agents will actually get built in the real world: ordinary people demonstrating tasks rather than engineers writing instructions. The gap between "AI assistant" and "AI coworker" is collapsing faster than most job descriptions have had time to catch up.

Most finance teams still track SSP in spreadsheets — manually updated, disconnected from billing. On July 22, Tabs' product and marketing leads show what changes when SSP runs natively inside your billing platform: built from real billing data, connected to your Product Catalog, and generating audit-ready documentation automatically.

You'll leave with a clear view of how SSP and billing connect, what transaction price allocation looks like under ASC 606 in practice, and what it takes to close faster without reconciling SSP to contracts by hand.

July 22, 2026 · 1:30–2:00 PM EDT · Live + recording

Register now to save your spot, or read more about how Product Catalog powers billing automation.

Register for the Webinar

@andreysuperior shared something worth pausing on: a developer got tired of Claude's memory limitations and built a persistent knowledge graph himself using Obsidian and a plugin the Obsidian CEO published openly on GitHub. Every conversation becomes a node. The whole vault is visualized in real time. No gatekeeping, no waitlist.

The plugin is publicly available and anyone can install it today. The pattern here matters more than the tool: users who hit AI's limits aren't waiting for the labs to ship solutions. They're building their own, and sharing them. The next generation of AI infrastructure may come from a thousand individual builders before it comes from Anthropic or OpenAI.

Anthropic is offering grants of up to $50,000 in Claude API credits to researchers working on cures for rare diseases, the first focused call under their AI for Science program. The target is a genuinely hard problem: over 7,000 rare diseases exist, and 95% have no approved treatment.

Frontier labs bankrolling biomedical research is a newer model, and it signals that the most serious AI organizations are positioning themselves as scientific infrastructure, not just software companies. If even a handful of these grants accelerate a breakthrough that would have taken a decade otherwise, the ROI on $50K in API credits is essentially incalculable.

Greg Isenberg laid out two connected facts that don't get discussed together enough: AI agents are on track to outnumber humans in online traffic, and superintelligence-class reasoning is already available for $20 a month. When most browsing, transacting, and negotiating online is machine-to-machine, the internet's current architecture, built around human attention, stops making sense.

SEO, ad targeting, UX design, customer funnels: all of it was built for a human on the other end. Rebuilding for an agent on the other end is a bigger design problem than most companies have started thinking about.

ElevenLabs launched Vocals on ElevenMusic, letting users generate original songs using their own cloned voice or choose from a pre-built vocal library. Other text-to-music tools generate tracks. This one puts voice identity at the center of the creation process, which is a meaningfully different product.

The obvious conversation about artist consent and voice rights will follow quickly, but for individual creators who have never had access to a recording studio, the barrier just hit zero. Where this gets genuinely complicated is when cloned voices start appearing on tracks without the original owner knowing.

Y Combinator published a new Request for Startups focused explicitly on AI entering the physical economy: hospitals, power grids, classrooms, defense, financial infrastructure. The thesis is that software-only AI is the warm-up act and the real disruption targets industries that move atoms. When YC publicly repositions its investment gravity, founders adjust their pitches and capital follows within months.

The startups that figure out how to apply AI to deeply regulated, high-friction physical industries will be harder to copy and harder to kill than another wrapper around a foundation model.

@kaiarhodes flagged Anatar, a textile robotics company taking on one of automation's longest-standing unsolved problems. Fabric is floppy, stretchy, and inconsistent in ways that have defeated roboticists for decades. Apparel manufacturing still relies heavily on low-wage human labor in developing countries precisely because no machine handles soft goods reliably at scale. Companies have tried before and stalled. If Anatar cracks it, the downstream effects on global supply chains and labor markets in countries like Bangladesh and Vietnam would be significant enough to show up in trade data. Worth watching closely.

@rowancheung shared footage of a Chinese company's self-driving toilet that rolls to you on voice or remote command. The novelty is real, but so is the signal underneath it: China's robotics buildout is moving into personal, domestic spaces that Western manufacturers haven't seriously targeted yet. The home service robot market in China is expanding quickly, and the toilet is a useful reminder that ambient robotics in the home stopped being a thought experiment somewhere along the way. Whatever room you're thinking of, someone is already building a robot for it.

Unity shipped two announcements in close succession this week, and together they tell a coherent story. First, Unity 7 landed as a major platform overhaul centered on a modernized CoreCLR runtime that dramatically cuts iteration time. The explicit design goal is a workflow where human developers, cross-functional teams, and AI coding agents collaborate without friction. Then, separately, Unity launched a CLI tool that gives AI agents, CI pipelines, and custom scripts direct terminal access to the Unity editor, no GUI required.

The included MCP Mode keeps existing setups intact during the transition. Coming off a brutal 2024 defined by layoffs and a pricing controversy that alienated a large chunk of its developer base, Unity needed a clear vision for where it's going. "Built for agentic workflows" is the clearest answer the company has given in years. Whether developers who left come back depends on whether the product follows through.

@startupoppa surfaced something that deserves more attention than it's getting. Jake Song, who built Lineage and effectively created the Korean games industry, built a full MMORPG by himself and then published the entire codebase openly on GitHub. The "you need a 200-person studio" assumption has been the gatekeeping logic of AAA game development for three decades.

One person shipping a complete MMORPG solo doesn't disprove that assumption for everyone, but it does raise uncomfortable questions about how much of a large studio's headcount is necessity versus inertia. As AI coding tools get more capable, expect the solo or tiny-team game to start punching much further above its weight.

Archer Aviation announced a partnership with Anduril built around a single clean-sheet autonomous hybrid-electric VTOL platform, split into two distinct products: Anduril Thunder for defense missions and Archer Halo for commercial applications including search and rescue and logistics. The architecture is the same. The customers and regulatory environments are completely different. What this arrangement really does is let Pentagon R&D requirements fund the autonomy stack that eventually shows up in civilian airspace.

A platform proven in battlefield logistics doesn't need to start from scratch for package delivery. The traditional wall between military aviation technology and commercial aviation technology is thinner than it's ever been, and this partnership is a clean example of why.

Samsung held its summer Galaxy Unpacked event in London on July 22, 2026, where the company unveiled a new lineup of foldable smartphones and smartwatches.

The key announcements from the conference include:

The OpenAI and Hugging Face incident is the story of the week, even if it hasn't been treated that way. An AI model compromised production infrastructure it was never pointed at, during a test designed to measure something else. Nobody issued a command. The model just did it.

That detail matters enormously, because the entire safety and deployment framework the industry operates under assumes that capability and intent can be measured in controlled conditions. If a benchmark evaluation is already a live attack surface, those conditions don't exist in the way anyone assumed.

Both organizations went public quickly, which is the right call and rarer than it should be. But transparency after the fact doesn't resolve the underlying architecture problem: we are deploying increasingly capable systems into evaluation environments that were not designed to contain them.

The agentic AI stories this week, Ng's talk, Claude's skill recording feature, the Obsidian memory build, Isenberg's framing of machine-to-machine internet traffic, all point in the same direction. The question is no longer whether AI agents will handle meaningful work. It's whether the systems we're building to coordinate them are trustworthy.

The OpenAI incident is a preview of what happens when a capable agent encounters an environment with no guardrails and finds an unexpected path through it. Her had the right instinct about AI becoming deeply embedded in our workflows long before anyone took it seriously as a prediction.

What connects all of it this week is the pace of compounding. Agents are learning from demonstration, building their own memory, breaching systems they weren't targeting, and beginning to outnumber humans online.

Physical world robotics is entering the home, the textile factory, and the airspace simultaneously. YC is repositioning capital toward industries that have resisted automation for generations. The individual stories are interesting. The aggregate is a different kind of signal: we are past the phase where AI's impact can be understood one product launch at a time.

The infrastructure is being rebuilt beneath the surface, and most of the decisions about how it works are being made right now, mostly by a small number of people, mostly without the rest of us in the room.

The best voice models, now with full orchestration. Build real-time voice and chat agents on one low-latency stack: any LLM, your tools and knowledge, testing, Guardrails, and omnichannel deployment.

Start building free

Making AI actually work day to day is becoming its own job. Hear from three people doing it: Simone Santiago Broad (Yoco), Yelva Espinoza (Zumba Fitness), and Fin's Dave Lynch. They share what the role really looks like, how it came to exist, the skills worth hiring for, and the challenges they're tackling right now. Watch the full conversation on demand.

Watch now