In partnership with
In Today’s Issue:
🛡️ OpenAI arms the defenders with GPT-5.5-Cyber
🇺🇸 Anthropic's Fable ban starts to thaw
🛰️ SpaceX rents Colossus to Reflection for $6.3B
🤖 Humanoids clock in at Automate 2026
🎬 A Notion agents deep-dive and Seedance 2.5
✨ And more AI goodness…
⚡ The Signal
The two biggest AI labs just gave opposite answers to the same question: what do you do with a model that can break software at scale?
OpenAI's answer is to ship it. Its Daybreak expansion released the full GPT-5.5-Cyber to vetted defenders and pulled more than 30 open-source projects into a "Patch the Planet" cleanup. Anthropic's answer came less than two weeks ago, when Washington forced the company to pull Fable 5 and Mythos 5 over the same offensive cyber capability. OpenAI gates access through verification and trust signals rather than a hard lockdown, wagering that handing defenders the better model first tilts the field their way. That bet now runs in production, and the rest of us are downstream of whether it holds.
All the best,
Kim Isenberg
(World Economic Forum)
🇺🇸 Anthropic's Fable Ban Starts to Thaw
Less than two weeks after Washington forced Anthropic to disable Claude Fable 5 and Mythos 5 worldwide on national-security grounds, the standoff is cooling rather than resolving. President Trump told Axios he no longer views Anthropic as a national-security threat after meeting CEO Dario Amodei at the G7, yet the Commerce Department's export-control directive is still legally in force and Fable remains dark. Anthropic's own fix may arrive first: an identity-verification system due July 8 could let it confirm US citizenship and quietly restore Fable to domestic users without the ban being formally lifted.
(Reuters)
🛰️ SpaceX Turns Colossus Into a $6.3B Compute Landlord
Open-source AI startup Reflection has agreed to pay SpaceX about $150 million a month for access to Nvidia GB300 chips at the Colossus 2 data center, a deal worth up to $6.3 billion through 2029. It is the latest sign that SpaceX has turned Colossus into a commercial compute platform, after leasing capacity to Anthropic (around $1.25 billion a month for 220,000-plus GPUs at Colossus 1) and Google (about $920 million a month for 110,000 GPUs). Compute is quietly becoming the asset that prints recurring revenue - and SpaceX possibly the next Oracle?
🎬 ByteDance Teases Seedance 2.5
ByteDance is teasing Seedance 2.5, the next version of its cinematic video generator, built around a director-style workflow of reference images, camera control, and shot pacing. The company has signaled a mid-2026 release targeting 4K output and near real-time generation, a step up from Seedance 2.0's 1080p clips of up to 15 seconds. Official specs are not out yet, so the demo reel is best read as a preview of where ByteDance wants to take controllable, on-brand video.
Use an AI model as your personal red team.
Why it helps: The same models now strong enough to find software vulnerabilities can audit your own digital exposure before someone else does, no security team required.
Try this: paste this into your assistant: "Act as a friendly security red-teamer. Ask me 8 quick questions about my accounts, devices, backups, and 2FA. Then give me a prioritized list of my five biggest security risks and the exact fix for each one I can do this week."
🎬 Watch This
Peter and I sat down with Akshay Kothari, co-founder of Notion, to get into how Notion is shifting from a tidy notes app into an agent-first workspace, and what that says about where software goes next. Kothari walks the arc from modular blocks to templates to today, where humans, custom code, and AI agents all share the same canvas.
The part especially worth the watch is "Smilers," Notion's internal people-ops agent. Ask it something and it answers from company knowledge; stump it and it opens a ticket, learns the answer, and writes it back, so the system quietly teaches itself. A company that repairs its own memory.
From there it is Notion Workers, model optionality, and Kothari's case that AI nudges companies to measure outcomes over headcount. His bet: the future of work is not one giant model doing everything, but millions of small, shareable agents spreading expertise across a team.
“The current systems are getting pretty good at cyber. You need to make sure that the defences are stronger than the offences.”
What looked like a throwaway line turned into a public dustup. After OpenClaw's Peter Steinberger pitched OpenClaw as a non-profit alternative to VC-funded rivals with "other agendas," Teknium, co-founder of Nous Research and a driving force behind the Hermes models, fired back. He appeared to take it as a swipe at Hermes, arguing that a non-profit label buys no real independence in AI, especially after OpenAI spent years turning the term into a political flashpoint. It is a small spat with a bigger tell: as open-source agents go mainstream, the fight is moving to governance, funding, and who actually controls the roadmap.
OpenAI Hands Its Best Cyber AI to the Defenders
The Takeaway
👉 OpenAI expanded Daybreak, its cybersecurity program, with the full GPT-5.5-Cyber model, a Codex Security plugin, and a new Cyber Partner Program.
👉 GPT-5.5-Cyber set a new state of the art on the CyberGym vulnerability benchmark at 85.6%, up from 81.8% for GPT-5.5, and even Anthropic's banned Mythos 5 (83.8%).
👉 A "Patch the Planet" push with Trail of Bits and HackerOne has signed up 30-plus open-source projects, including cURL, Go, and Python.
👉 Access stays gated to "trusted defenders" through verification, OpenAI's open-but-controlled answer to Anthropic's locked-down Mythos.
OpenAI is betting that the safest thing to do with a top-tier offensive cyber model is to hand it to the defenders first. On June 22 the company expanded Daybreak, its security initiative, and released the full version of GPT-5.5-Cyber through a limited program for vetted defenders. The model finds, validates, and helps patch software vulnerabilities, and it now leads the CyberGym benchmark, a test of automated vulnerability discovery and exploitation, at 85.6%, ahead of the 81.8% posted by the general-purpose GPT-5.5, and it even clears Anthropic's now-banned Mythos 5 (83.8%).
(OpenAI)
The release is more than a model: A new Codex Security plugin pulls vulnerability detection and patch validation straight into the coding workflow, while the Daybreak Cyber Partner Program lets vendors like Sophos wire OpenAI's models into their own products. Alongside it, OpenAI launched "Patch the Planet," an effort with Trail of Bits, HackerOne, and project maintainers to move widely used open-source software from findings to fixes. More than 30 projects, among them cURL, Go, Python, Sigstore, and pyca/cryptography, have signed on.
(OpenAI)
The strategy is deliberately the opposite of locking the technology away. "AI is already good and about to get super good at cybersecurity," CEO Sam Altman said. "We'd like to start working with as many companies as possible now to help them continuously secure themselves." OpenAI argues it is "not practical or appropriate to centrally decide who gets to defend themselves," and gates access through verification and trust signals instead. That lands less than two weeks after Washington forced Anthropic to pull Mythos 5 and Fable 5 over the same kind of offensive cyber capability.
Why it matters: The frontier labs are splitting on how to handle AI that can break software at scale. Anthropic waited for the government and switched its models off; OpenAI is shipping its sharpest cyber model to as many verified defenders as it can, betting that defense will outrun offense. The open question now: having forced Anthropic to pull the weaker Mythos, will Washington let OpenAI keep shipping the stronger GPT-5.5-Cyber, or come knocking next?
Sources:
🔗 https://openai.com/index/daybreak-securing-the-world/
🔗 https://www.axios.com/2026/06/22/openai-rolls-out-more-capable-version-of-cyber-model
The IT strategy every team needs for 2026
2026 will redefine IT as a strategic driver of global growth. Automation, AI-driven support, unified platforms, and zero-trust security are becoming standard, especially for distributed teams. This toolkit helps IT and HR leaders assess readiness, define goals, and build a scalable, audit-ready IT strategy for the year ahead. Learn what’s changing and how to prepare.
The chart: Artificial Analysis's GDPval-AA v2 leaderboard scores models with an Elo rating on real-world work tasks, anchored to a human baseline of 1,000. Claude Fable 5 tops it at 1,783, with Claude Opus 4.8 at 1,615. The highlighted bar is the open-weights GLM-5.2 at 1,524, which edges GPT-5.5 (1,509) for third place. Everything down to Grok 4.3 (1,094) still clears the human line.
The lesson: Open-weights models are no longer just chasing the frontier on coding. A freely available Chinese model now beats GPT-5.5 on general real-world work and trails only Anthropic's two flagships.
The caveat: Elo on a fixed task set is not the same as economic value, and the top scores lean on maximum reasoning settings ("max," "xhigh," "with fallback") that cost more time and money than everyday use.
🤖 Humanoids Clock In at Automate 2026
⚡ Bottom line: Automate 2026 opened in Chicago this week with humanoid robots that already hold paying jobs on real factory and warehouse floors.
💡 Why it matters: The bar for humanoids has shifted from walking demos to signed deployment contracts and units shipped.
🔎 What it means: Physical AI is turning into an industry, pairing robot foundation models with mass-produced bodies and live customers.
Automate 2026, the big automation show running June 22 to 25 at Chicago's McCormick Place, put humanoid robots at the center for the first time. The Association for Advancing Automation gave the category its own NVIDIA-sponsored Humanoid Robot Pavilion, and "physical AI" ran through the keynotes. The shift on display: these robots are no longer auditioning. Boston Dynamics has started commercial shipments of its electric Atlas, with its entire 2026 run reportedly committed to Hyundai and Google DeepMind.
(Boston Dynamics)
The clearest proof is in the aisles. Agility Robotics' Digit is working paid shifts at a Toyota plant in Woodstock, Ontario under a robots-as-a-service deal, and has already moved more than 100,000 totes at a GXO warehouse. The AI underneath is what changed: vision-language-action models like NVIDIA's Isaac GR00T let one robot follow plain-language instructions and handle multi-step tasks, so a single platform can be retrained for many jobs instead of hand-coded for one. The bodies are finally catching up to the brains.
(Agility Robotics)
10x the context. Half the time.
Speak your prompts into ChatGPT or Claude and get detailed, paste-ready input that actually gives you useful output. Wispr Flow captures what you'd cut when typing. Free on Mac, Windows, and iPhone.
