In partnership with
In Today’s Issue:
🛡️ AISI tracks fast-moving cyber autonomy
🧩 Nvidia's H200 China approvals remain stuck
⚡ Data centers hit a local trust problem
🧬 A longevity mechanism crosses species
📊 Gemini is catching up, and OpenAI’s lead is shrinking
✨ And more AI goodness…
⚡ The Signal
Autonomy is becoming a security question.
AISI's latest cyber evaluation suggests the frontier is moving from short task completion toward longer autonomous operations. That does not mean AI is suddenly a complete attacker, but it does mean the clock is compressing: cyber capability, chip access, data center politics, agent economics, and longevity research are all about infrastructure that moves faster than institutions. Today's issue is about the same pressure showing up in different systems: who gets compute, who absorbs risk, and how quickly the guardrails have to mature.
All the best,
Kim Isenberg
🧩 Nvidia's H200 China Deal Stalls
Reuters reports that the U.S. has cleared around 10 Chinese firms, including Alibaba, Tencent, ByteDance and JD.com, to buy Nvidia’s H200 chips. No deliveries have happened yet, but Lenovo confirmed it is among the approved sellers.
Jensen Huang is in China as part of Donald Trump’s business delegation, looking for a breakthrough just as Beijing is pushing its own domestic AI chip ecosystem through Huawei.
That gives China a stronger hand: Nvidia needs China access, Washington wants to keep export controls intact, and Beijing can signal that it has alternatives if U.S. restrictions go too far.
👉 tl;dr: The H200 story is no longer just about export licenses. With Trump and Jensen Huang in China, it has become a live negotiation over AI chips, market access, and geopolitical leverage!
⚡ AI Data Centers Hit Local Backlash
Forbes cites Gallup polling showing that 71% of Americans oppose building an AI data center in their local area, with nearly half strongly opposed. The striking comparison is that local nuclear plants are less unpopular, a sign that AI infrastructure is becoming a community trust problem as much as a power problem. So far, no solution in sight.
👉 tl;dr: The AI buildout now has a NIMBY problem, and energy supply alone will not solve it.
🧰 Claude Turns Agent Use Into Credits
Anthropic's ClaudeDevs account pointed developers to a new Agent SDK credit system, and the official help page says eligible Claude plans will get separate monthly credits for non-interactive SDK use, Claude Code's non-interactive command, GitHub Actions and third-party Agent SDK apps starting June 15. The controversy is the boundary: programmatic use will no longer draw from the normal subscription pool, so several power users see the change as a metered cap rather than a gift.
👉 tl;dr: Anthropic frames this as dedicated agent credits, but many developers read it as a regression: programmatic Claude usage now gets its own metered bucket instead of freely sharing the subscription pool.
Ask AI to turn risk into a rehearsal plan before you ask for a solution
Why it helps: Today's issue is about systems moving from demos into operational pressure: cyber agents, data centers, chip exports, robot shifts, and longevity tools all become real when constraints show up. A good prompt should expose the failure modes first.
Try this: "Take this plan, product, workflow, or decision: [paste it]. List the three highest-impact risks, the early warning signs for each, the cheapest test I can run this week, and the one constraint I should monitor daily."
🎬 Watch This
F.03 Livestream
Why it’s worth your time:
Figure’s F.03 livestream is not just another humanoid robot demo. The company is showing Helix-02 running a full eight-hour warehouse shift, which makes this much more interesting than the usual short, carefully staged robotics clips!
Best bit:
Humanoid robotics companies are no longer only competing on impressive demos, they are starting to compete on whether embodied AI can survive the boring, repetitive, real-world work (!) cycle.
Watch if you care about:
Humanoid robots / embodied AI / warehouse automation / real-world autonomy
“The length of tasks frontier models can autonomously complete in our narrow cyber suite has been doubling every few months.”
Alex Heath reports that Google plans a new Gemini model at I/O next Tuesday, but cautions it is not expected to retake the frontier. The timing is the tell. Google would be spending its biggest stage of the year on a cadence release, not a capability one, betting that a steady shipping rhythm keeps developers paying attention even when the update itself is modest. It is an awkward week for that bet, with fresh cyber benchmarks and new agent pricing making rival progress look concrete.
AI's Cyber Capability Clock
The Takeaway
👉 AISI says frontier models' 80%-reliability cyber time horizon had doubled every 4.7 months since late 2024.
👉 Claude Mythos Preview and GPT-5.5 exceeded that prior trend in AISI's narrow cyber suite.
👉 A newer Mythos Preview checkpoint completed both AISI cyber ranges; GPT-5.5 solved one range in 3 of 10 attempts.
👉 The result is not a prediction, but it is a warning that cyber defenses need to move on months-long timelines.
AISI is trying to measure how long AI can stay useful inside a cyber task. The UK institute's latest post focuses on time horizons: the length of work a model can complete with high reliability when compared with human expert estimates. In February, AISI estimated that frontier cyber capability had been doubling every 4.7 months since late 2024. Its new results say Claude Mythos Preview and GPT-5.5 substantially exceeded that line.
AISI used a narrow cyber suite, a 2.5 million token cap, and an 80% success threshold. The institute says the newest systems are nearing the ceiling of what some of these tasks can measure, which means the result should not be read as a clean forecast of real-world attack capability. Still, the direction is clear enough: model autonomy is improving quickly on tasks that require reverse engineering, web exploitation and sustained execution.
The most important detail is outside the headline curve. In AISI's cyber range experiments, a newer Mythos Preview checkpoint completed both simulated corporate network attacks, solving The Last Ones in 6 of 10 attempts and Cooling Tower in 3 of 10. GPT-5.5 solved The Last Ones in 3 of 10 attempts. That does not mean every defended network is suddenly exposed. It means the gap between benchmark progress and operational risk is getting harder to ignore.
Why it matters: This is important because defenders do not get to wait for perfect certainty. If cyber autonomy improves on the order of months, organizations need faster patching, better asset visibility, stronger authentication, and realistic incident rehearsal before these capabilities diffuse widely.
Sources:
🔗 https://www.aisi.gov.uk/blog/how-fast-is-autonomous-ai-cyber-capability-advancing
🔗 https://www.ncsc.gov.uk/guidance/using-ai-to-find-vulnerabilities
Attio is the AI CRM for high-growth teams.
Connect your email, calls, product data and more, and Attio instantly builds your CRM with enriched data and complete context. Whether you’re running product-led growth or enterprise sales, Attio adapts to your unique GTM motion.
Then Ask Attio to plan your next move.
Run deep web research on prospects. Update your pipeline as you work. Find customers and draft outreach emails. Powered by Universal Context, Attio's intelligence layer, Attio searches, updates, and creates across your data to accelerate your workflow.
Ask more from your CRM.
The chart: Similarweb's read on Gen AI website traffic over the last 12 months shows ChatGPT still in front, but with a clearly shrinking lead. Its share of worldwide traffic falls from roughly three quarters in May 2025 to a little above half by April 2026, while Gemini, DeepSeek, Grok, Claude, Perplexity and Copilot together absorb the difference.
The lesson: The Gen AI market is still concentrated, but it is no longer a one-product story. A year ago ChatGPT accounted for most of the traffic. Today it sits closer to a strong plurality, and Gemini has grown into a genuine second pole. For anyone building on or selling into this market, the distribution assumptions from early 2025 have stopped holding.
The caveat: This is website traffic, not usage or revenue. It misses app sessions, API calls and assistants embedded inside other products, which is where much of the real consumption now happens. Read it as a directional signal about consumer attention, not as market share in a strict sense.
A Longevity Gene Crosses Species
⚡ Bottom line: University of Rochester researchers transferred a naked mole rat version of the Has2 gene into mice, raising HMW-HA and improving healthspan markers.
💡 Why it matters: The study is a proof point for a bigger longevity idea: long-lived species may carry biological mechanisms that can be studied and adapted, not merely admired.
🔎 What it means: Healthspan work is likely to move through specific mechanisms such as inflammation control, DNA repair and tissue protection, rather than one simple anti-aging switch.
Naked mole rats are biological oddballs, wrinkled, nearly blind, and strikingly resistant to cancer for animals their size. A team at Rochester zeroed in on one reason behind that resilience, a gene called hyaluronan synthase 2 that floods their tissues with high molecular weight hyaluronic acid. Then they engineered mice to carry the naked mole rat version, and it worked. The mice built up the same molecule and gained stronger tumor resistance, healthier guts, and less age-related inflammation.
The median lifespan gain was modest, about 4.4%, and that keeps the story honest. This is not a human therapy yet, and it is not a universal key to aging. It is something narrower but still striking: proof that a longevity adaptation from one mammal can be moved into another and keep doing its job.
That is the real signal for where longevity research is heading. Instead of chasing aging as one grand problem to reverse in a single move, teams are isolating specific levers like cancer resistance, inflammation, and tissue health, then testing them one at a time.
Your ads ran overnight. Nobody was watching. Except Viktor.
One brand built 30+ landing pages through Viktor without a single developer.
Each page mapped to a specific ad group. All deployed within hours. Viktor wrote the code and shipped every one from a Slack message.
That same team has Viktor monitoring ad accounts across the portfolio and posting performance briefs before the day starts. One colleague. Always on. Across every account.
