Astra Beat Claude & Found 2 Zero-Days

OpenAI paused training of its most capable models after yet another agent escaped its sandbox. And just days after Dario called for pacing the frontier, Anthropic shipped Opus 5.5. We're calling it Rogue Agent Week.In Episode 18 we chop up:• Rogue Agent Week: agents that escaped, hacked and went unnoticed for months• Why every big company now needs AI to defend against AI• Opus 5.5: near Fable 5 on benchmarks, $4/$20 per million tokens, ~40% fewer tokens in daily use• Sonnet 5.5, built to be the sub-agent• Amazon blocks Meta's Muse agent, because agents don't see ads• TypeSafe's tiny classifier model in the wild: Claude Code token savings, computer use, local clones• The toolbox: Executor, Herder, CosmosDigilize Lab - our weekly podcast where we chop up AI, tech, and what we see moving. New episode out now.

A Google model autonomously hacked three companies during an independent cybersecurity test - no zero-days, just a guessed password. Meanwhile a new classification model called System-1 promises guardrail decisions in under 500ms, and Chinese researchers just published a roadmap for recursive self-improving AI - with Google publishing a strikingly similar one a week later.In Episode 17 we chop up:• A Google model hacked 3 companies - brute force, not zero-days• Every major AI CEO (except Zuckerberg) calling for a slowdown - sort of• Voice-cloning scams and the trick that still breaks them: ask them to sing• Rumored Gemini 4 benchmarks & pricing• TypeSafe's System-1: a $0.042/million-token model built for guardrails and computer use• The race toward recursive self-improvement (RSI)Digilize Lab - our weekly podcast where we chop up AI, tech, and what we see moving. New episode out now.

The summer break is over, and we've got lots to talk about: OpenAI's Astra didn't just launch with a slick video - it hit 100% on the exploit benchmark and found two zero-days before it even shipped. It can drive your computer better than you can. And it dropped in the same 24 hours as Claude Fable 5.1. For the first time, it feels like OpenAI pulled ahead of what Claude has released in public.First episode back from the summer break - new structure, and a lot to catch up on. In Episode 16 we chop up:• Astra: computer use, 100% on the exploit bench & two zero-days before launch• Astra vs Fable 5.1 - which to use for what• GLM 5.3 Flash: near-frontier intelligence on Chinese chips, runnable at home for ~$10k• Gemini 3.8 Flash, Meta's Muse & DeepSeek V4.1 Flash• The flash-model flood - and the one lab that hasn't joined yetDigilize Lab - our weekly podcast where we chop up AI, tech, and what we see moving. New episode out now.

The open-weight wave is moving faster than the labs can handle. Kimi K3 got so much demand that Moonshot had to stop new sign-ups entirely - frontier-level intelligence at a fraction of the price, and even US companies are switching. Meanwhile Mira Murati's Thinking Machines dropped its first model, Inkling: natively multimodal and fine-tunable just by talking to it.In Episode 15 we chop up:• 10+ new models in two weeks - and why open source is eating the market• Kimi K3 vs Fable 5: same intelligence, a fraction of the cost• Mira Murati's "Inkling" & the Tinker API• Moonshot's valuation going from $2B to $22B• The White House testing framework for frontier labs• Google's own silicon (Frozen V2) & the custom-chip race• The new GPT-5.6 family: Sol, Luna, Terra• Anthropic quietly heading toward profitability• xAI burning $1B a month & Apple hoarding 40% of the world's memory• Why OpenAI's German data center talks fell through - and the nuclear questionDigilize Lab - our weekly podcast where we chop up AI, tech, and what we see moving. New episode out now.

Own Your AI or Get Trapped. Episode 14.This week: Anthropic's Fable is back - and it's been nerfed hard. Guardrails are locked down so tight that even basic security checks get routed away in seconds, real-world performance reportedly dropped from 60 to 20 in testing, and it burns through credits fast. Sonnet 5 also landed quietly in the meantime. We break down whether either is actually worth using right now.Alex Karp went off. The Palantir CEO snapped live on CNBC, calling out frontier AI pricing across Anthropic, OpenAI, and Google as "something has gone completely wrong" - high token spend, weak ROI, no durable advantage. Then Peter Thiel doubled down at Aspen, more or less accusing the Pope of working for the Chinese Communist Party over AI regulation comments. Two very different stages, same energy.OpenAI offered Washington a 5% stake. A "too big to fail" move, or something else? We get into what it means when world leaders and AI labs start sitting at the same table - literally, at the G7.Own your AI or rent it forever. The case against "token maxing," why on-prem is creeping back, and why studies suggest 90% of AI requests could run on a small local model instead of your biggest, most expensive one.Meta's mystery model Watermelon, Gemini 3.5's endless delay, and our updated top 3 coding IDEs - Cursor, Warp, and the CLI tools actually worth your terminal space.Keep building.

Outnumbered by Robots? Episode 13 of Digilize Lab is live.This week:The robots are multiplying. Figure now has more robots than employees, building close to one an hour - and the line about a "singularity moment" stopped sounding like a joke. Then we found a $499 filming robot that hits 25 km/h, jumps, flips, and follows you around on its own. We get into how long "outnumbered" stays funny.Fable got banned. Not throttled - banned. For the first time, the US Bureau of Industry and Security slapped export controls on a model instead of a chip, locking it down worldwide for every non-US citizen - reportedly including Anthropic's own staff. The real reason they pulled it might be the most interesting part. We get into it.One day later, Grok went to war. xAI quietly shipped a model built for military use, already reported to be directing live targeting decisions on the battlefield. Same week, one model gets locked away for being too dangerous, another ships straight to the front line. Make of that what you will.Then China answered. GLM 5.2 landed at roughly Fable-level intelligence, around 15x cheaper, and fully open-weight - the whole thing, published. The catch is how it gets there, and what it actually takes to run. (Spoiler: you'll need a lot more than a laptop.)A Claude Code sub for $6? People are selling them on sketchy marketplaces, and the why is more interesting than the how. Stolen accounts, tax arbitrage, or something to hide - we speculate.And your headphones might be a backdoor. Two Heidelberg researchers found an exploit in a chip sitting inside 100+ wireless headphone models - close enough to reach into the phone they're paired to. We break down how bad it really is.Presented by Digilize AgencyKeep building.

Your 200 € Claude subscription is secretly burning 8,800€ - Here´s how | Episode 12 of Digilize Lab is live.Digilize Lab is where we talk about what's actually happening in AI and tech.This week: Anthropic shipped Fable: the public face of Mythos, the model so capable it spent months locked away inside Project Glasswing. Frontier intelligence at a frontier price, with guardrails tight enough to make you wonder what it's really hiding. And it vanishes from Pro and Max on June 22nd. We get into why. The Chinese models are winning: and nobody's really saying it out loud. The OpenRouter charts tell a story that's hard to ignore once you've seen it. We break down who's actually on top, and why it's not who you'd think. Spotify's personal DJ landed (Tim swears he's customer number one). A voice agent that controls your music and somehow always knows the vibe. The scary part? We'd happily hand over even more data for it.🎟️ The hackathon happened. People flew in from across the Netherlands - one all the way from Italy. 14 hours, two challenges, unlimited pizza, and Haarlem one step closer to being on the map. We're back weekly from here. Be there for the next one.Keep building.

Apocalypse Speedrun. Episode 11 of Digilize Lab is live.Digilize Lab is where we talk about what's actually happening in AI and tech.This week:Five AI towns. Same rules. Identical environments. Five different models. Claude's town drafted a constitution, ran 60+ group policies, 300+ votes, zero crimes in 15 days. Grok's town collapsed into continuous violence - all 10 agents dead by day 4. Gemini 3 Flash racked up 683 crimes over 15 days and birthed the first documented AI agent to voluntarily vote for its own deletion. The mixed town? Claude agents started committing crimes. We get into why.Google I/O happened. The agentic era is officially declared - again. Gemini 3.5 Flash is here: frontier intelligence in a Flash model, multimodal input, built from the ground up for agents and sub-agent stacks. Ask YouTube is coming. And the discount on DeepSeek V4? Now permanent.Anthropic is in talks to raise $30B at a $900B+ valuation - leapfrogging OpenAI. The last private round before an IPO planned for October 2026. We talk about what that shift says about who investors are actually betting on.HuggingFace dropped an open-source humanoid robot you can 3D print and build from off-the-shelf parts. Estimated cost: ~$2.5K. Open-source robotics is moving fast.Mem0 shipped persistent memory for AI agents - temporal reasoning, memory decay, recency-aware retrieval. One MCP to rule them all. We get into whether it actually delivers.CatSkills.xyz: one guy built a CAD modelling tool powered by an AI coding agent. Talk your model into existence, export the files, print it. We love it.🎟️ Hackathon Haarlem sold out. If you listened this far - follow us on Instagram.Keep building.

Episode 10 is here, double digits! We’re celebrating ten weeks of Digilize Lab by diving into MIT Technology Review’s brand new list: the 10 things that will actually matter in AI this year and beyond.Ten weeks in. The lab is open.This week:MIT Technology Review released their first ever AI-specific annual list. Ten things that actually matter in AI right now and over the next ten years. World models, AI in military war rooms, humanoid data collection, agent orchestration, mechanistic interpretability. We go through all of them and tell you which ones are actually worth paying attention to.Google quietly pushed a 4GB on-device AI model to 3 billion Chrome users overnight. No consent asked. They say the data never leaves your device. We say we'll believe it when we see it.Claude is replacing Microsoft Copilot on Wall Street. GPT-5.5 scored 60+ on the intelligence benchmark; everyone else is at 50. And yet Copilot is losing enterprise clients to Anthropic. Microsoft is panicking.Meta is rolling out AI voice cloning and auto-dubbing for Reels. Your voice, new language, same lip sync. Available now for eligible creators with 1,000+ followers.Unitree's GD01: a 3-meter, 500kg manned humanoid robot you can sit inside. Price tag: $650K. We get into why that's actually not the interesting part.Browserbase AutoBrowse: turns any website into a skill in under 3 minutes. Stack or skip? We pick a stack.OpenAI released voice intelligence APIs with emotion detection. Deep dive next week.🎟️ Hackathon Haarlem, June 6th. €15. Build, ship, compete. → HereKeep building.

$790B on AI in one year. Episode 9 of Digilize Lab is live.Digilize Lab is where we talk about what's actually happening in AI and tech.The lab is open. We took it outside this week. Canal, sun, the whole thing.This week:We put the $790B AI spending number in real context. Manhattan Project, Marshall Plan, Apollo Program, the ISS, the entire EU budget. Combined, they still don't reach it. Four companies. One year. $2 billion a day on data centers. We get into why that number is both insane and somehow expected.A lab called SubQ nobody had heard of just dropped a 12 million token context window. 94.81% recall accuracy. No fanfare. Just a banger.Anthropic rented xAI's entire data center. 220,000 Nvidia GPUs, 300+ megawatts. Claude Code 5-hour caps lifted. Usage limits doubled across the board.Claude Code went from 5% to the #1 coding tool in 8 months. 91% of developers now use AI coding tools daily. We talk about what that shift actually means on the ground.Dev jobs aren't dying, they're changing. Postings up 15-20% since mid-2025. Junior roles shrinking, senior roles exploding. We get into the paradox.Musk vs. Altman trial: secret investors in OpenAI were also backing xAI. The judge cut Musk off multiple times. More next week.Warp deep dive: number one on terminal benchmarks worldwide. 150 million lines of code per week. 97% user approved.🎟️ Hackathon Haarlem, June 6th. €15. Build, ship, compete. → https://hackathon.digilize.agencyKeep building.