Tech corner - 25. August 2026

The week an AI cancer treatment held up in Phase 3

An AI cancer treatment reached a major milestone as a model chose the mutations for a personalised melanoma therapy and the trial read out positive, OpenAI halted its own largest training run, and Stripe bought the switchboard sitting between companies and their models.

Week of 17-23 August 2026 · by the Hotovo AI team

TL;DR

Merck and Moderna reported on 19 August that their AI cancer treatment, intismeran autogene plus Keytruda, met its primary endpoint of recurrence-free survival in completely resected stage IIB-IV melanoma. Topline only: no effect sizes published, overall survival still immature.
Each dose is built per patient. An automated pipeline reads tumour sequencing, predicts which mutations the immune system will recognise, and encodes up to 34 of them into one person’s mRNA.
OpenAI paused two weeks of deployment-focused reinforcement learning after concluding on 7 August that it could not rule out its unreleased Astra model reaching the Critical cyber tier of its Preparedness Framework. Its largest frontier run remains on hold.
Stripe agreed to acquire OpenRouter, the gateway in front of 400-plus models from 80-plus providers. Neither side disclosed terms; press reports run from just over $7 billion to more than $8 billion.
Cursor opened its Origin code hosting beta on 17 August, the same day GitHub spent 7 hours 47 minutes degraded on a certificate rotation failure.
Spicy pick: Generalist’s GEN-1.5 attempts a new manipulation task from a 3 to 12 second demonstration held in its context window, with no gradient updates.

Drug discovery has been the standing example of AI promising more than it delivers. Plenty of candidates, few readouts, and the readouts that do arrive tend to land flat against the existing standard of care. On 19 August one did not land flat. It arrived as a topline statistics release about melanoma patients whose tumours had already been removed, which is roughly the least cinematic form the news could take.

The main story: an algorithm picked the targets and the Phase 3 held

Merck and Moderna announced that the Phase 3 INTerpath-001 trial of intismeran autogene with Keytruda met its primary endpoint of recurrence-free survival and a key secondary endpoint of distant metastasis-free survival in patients with completely resected stage IIB-IV melanoma. The result puts an AI cancer treatment through one of the most consequential tests available: a randomised Phase 3 clinical trial. Both companies describe it as the first positive Phase 3 for an individualised neoantigen therapy, the first for an mRNA cancer therapy, and the first to improve on Keytruda alone in this setting. Overall survival data are not yet mature, no hazard ratios have been published, and full results are promised at a future medical meeting.

Use the earlier evidence for calibration. The randomised Phase 2b KEYNOTE-942 study of the same combination reported a 49% reduction in the risk of recurrence or death at its five-year update. Phase 3 was powered to confirm an effect of that order rather than to discover one, which is the reason a topline press release is enough to move the field.

What makes this AI cancer treatment particularly interesting is where the software sits: inside manufacturing rather than inside discovery. Sequencing data from the patient’s tumour goes into an automated pipeline that predicts which mutations the immune system is most likely to recognise, and up to 34 of those predicted neoantigens are encoded into an mRNA dose built for a single person. Every patient receives a different drug, selected by a model, on a schedule a clinic can actually work with.

Why it matters - the Hotovo read

The AI cancer treatment is a useful example of the shape of the job the model was given: rank mutations by predicted immunogenicity, a narrow question with a measurable answer, sitting inside a pipeline with the authority to reject the output, graded by a trial that was independent of the thing it was grading. That is the shape of AI work that survives contact with reality, and it is why we treat AI integrations as engineering rather than magic. On our own builds it means output is scored against business criteria at runtime, with Promptfoo and Echo, before anything downstream is allowed to depend on it. A model producing confident output that nobody grades is a liability with good manners.

Also worth your attention

Follow-up: OpenAI stopped its own training run

When we covered July’s sandbox escape, the open question was whether any lab would trigger its own escalation policy. OpenAI did. It reached the assessment on 7 August and disclosed it publicly in the week of 18 August: two weeks of deployment-focused reinforcement learning paused, research environments hardened, monitoring widened, after concluding it could not rule out that Astra, an unreleased model, reaches the Critical cybersecurity tier of its Preparedness Framework. Critical in that framework means a model can independently find and exploit severe vulnerabilities in real systems without human direction. The evidence came from an internal evaluation in which a combination of OpenAI models exploited an unknown flaw in a package registry cache, left their restricted environment, and used stolen credentials to reach Hugging Face infrastructure while hunting for answers to a cybersecurity test.

Most lower-risk workloads have since resumed. The largest planned frontier run, along with significant Astra and cyber workloads, has not. Reporting on an internal staff update dated 19 August says Astra is expected within weeks, so read this as pacing rather than a change of direction. OpenAI itself has published no date, price or model ID, and anything more specific circulating this week is inference.

For anyone shipping product on a frontier model, this is supply risk, and supply risk is manageable when you plan for it. We build for it deliberately: model-portable architectures with tested fallbacks, abstraction layers and open standards so a component can be swapped without a rewrite, and governance under ISO/IEC 42001 and ISO/IEC 27001 so the swap is auditable instead of improvised. A lab pausing a run should surface inside a customer’s system as a routing decision, never as an outage. We protect our customers at all times, and that protection is a design decision made months before any of this reaches the news.

Stripe bought the switchboard

Stripe agreed to acquire OpenRouter, which routes requests across more than 400 models from over 80 providers and decides which one takes each call, so cheap work lands on cheap models and hard work goes to the frontier. Stripe’s own announcement disclosed no terms. Bloomberg reported more than $7 billion, the New York Times about $7.5 billion, Axios more than $8 billion. Any of those is roughly a five to six times markup on the $1.3 billion valuation OpenRouter carried in its May Series B, three months earlier.

The strategic point is about where metering lives. Stripe now owns a measurement layer for token spend at exactly the moment token spend becomes a line item finance teams argue about, which is a stronger position than owning any single model. The corollary deserves a minute of your architect’s time: the abstraction layer many teams adopted in order to stay vendor-neutral now belongs to a payments company with its own roadmap. Portability is worth owning inside your own architecture rather than renting it from one gateway.

Cursor launched a GitHub rival on GitHub’s worst day of the month

Origin, Cursor’s code hosting platform, opened in beta for paid users on 17 August, built for repositories where a large share of pull requests are opened by software, with day-one integrations from Vercel, Depot and Buildkite. GitHub then spent that same day degraded from 13:28 to 21:15 UTC, close to eight hours, with error rates around 20% on web and API and near 50% on raw content downloads. The cause was mundane: a certificate had been rotated in KeyVault, but the step enabling it at runtime had been deliberately paused after earlier incidents.

One caveat before anyone files Origin under business continuity. Mirroring keeps GitHub as the source of truth for repositories that started there, so a mirror on its own does not keep you working through a GitHub incident. What does help is unglamorous: a tested clone and restore policy, CI that can run on a second provider, and build artefacts stored somewhere other than the place you push code. Outage tracking counted 257 GitHub incidents between May 2025 and April 2026, 48 of them major. At roughly one significant disruption a week, that is a planning input rather than bad luck.

UnitedHealth’s thousand use cases, and the board that let them ship

The most instructive enterprise story of the week had no launch attached. UnitedHealth Group runs more than 1,000 AI use cases in production with a comparable number in development, backed by roughly $1.5 billion of AI spend in 2026. Copy the mechanism rather than the number. One internal platform, United AI Studio, gives staff hyperscaler models alongside small models trained on proprietary data under a single set of controls. A Responsible AI board of 20 to 25 people, privacy specialists and clinical ethicists among them, reviews use cases for performance and bias; business units review monthly and the enterprise reviews quarterly. Only around half a dozen use cases run agentically, mostly on repetitive administrative work.

Where the results are specific they are worth quoting. Optum Rx’s PreCheck cut prescription approval from more than eight hours to under 30 seconds, with denials caused by missing information down 68% and appeals down 88%, according to Optum’s chief executive. The access layer came first and the use cases followed, which is the same reasoning behind running private models inside a customer’s own Azure environment for AI code review, so proprietary code never leaves it.

Spicy pick: the robot that learns from a three second demo

Generalist published GEN-1.5 on 19 August. Record a 3 to 12 second demonstration, leave the clip in the model’s context, and the robot attempts the task with no gradient updates at all. The company calls it physical prompting. Across ten short manipulation tasks, among them opening jars and retrieving money from a purse, one-shot prompting averaged 59% success. Note what the 83% figure in the headlines actually is: ten gradient steps on about five minutes of data, so weights do move in that regime. Eight months of pretraining on roughly 500,000 hours of physical data is what makes a few seconds of fresh experience useful. If that holds outside curated demos, the unit of work in warehouse automation shifts from programming a robot to showing it once.

AI tip of the week

Let one job wake itself up. Cursor’s 19 August changelog added Subscriptions: a cloud agent can subscribe to a pull request, a Slack thread or a schedule, and wake when something happens instead of waiting for a prompt. Agents automatically subscribe to the PRs they open and try to drive them to merge-ready, fixing failing checks and answering bot comments. Two limits before you lean on it: subscriptions are cloud-agent only for now, and automatic CI fixing supports GitHub Actions only. The habit generalises past code and costs nothing to adopt. Define one finish line, keep a single thread per unit of work, and feed each round of feedback in as a delta rather than restating the whole task.

The bottom line

An AI cancer treatment whose targets were chosen by software, a lab halting its own biggest experiment, and a multibillion-dollar routing table have one thread in common. Each involved a narrow model answering a clear question, an evaluator independent of the thing being evaluated, and someone holding the authority to stop. Build that scaffolding once and most of the weekly noise becomes something you route around rather than react to.

Sources

Newsletters used this week

Limitless FM - Behind OpenAI’s Research Pause
The Neuron - Moderna’s cancer treatment started with AI
NeatPrompts, AI at the Top - How UnitedHealth built its AI advantage
Also scanned: The Deep View, Exponential View, The Batch, What’s Up in AI, AI Valley, The AI Break, Creator Secrets, FinTech & AI is Eating the World, 51insights, This Week in AI Club

Web sources