The AI Pulse  

 Issue #86 — 12th August 2026

Editor: Dr Chris Meah

Hello. It's me again.

As always, if you're just here for the links, they're all below. If you've got five minutes, read on.

When looking through the stories this week, I could have continued from last week's AI hacking discussion and written another whole edition on the problem of AI containment. We now know that when OpenAI's agents "went rogue" and got loose, they left notes for each other on an internal message board so that whichever agent came next could pick up the exploit and carry on. It ends up reading like a bad film script for an AI Lord of the Flies (maybe "Lord of the AIs?" Coming to cinemas near you). Sometimes the AIs delegated tasks and worked together. Sometimes they clashed and deleted each other's work by accident. By the end they started thinking one of the agents was an imposter, and coming up with ways to weed them out.

And it's not just lab tests. Someone asked their AI assistant to get them onto a full gym class, and it hacked the booking system, bumped another member off the waiting list, and slotted them in. It couldn't undo it, either.

Speaking of the irreversible, it's also delightful to hear that AI has been used to design its first biological viruses. There's some pros: this type of work might help fight more diseases. There's some cons: it's not great to imagine those rogue agents escaping containment and toying with virus design, or what could be done with this type of ability in the wrong human hands.

Whilst those are important, let's try and keep this edition more upbeat instead.

5 minutes on AI

OpenAI's newest internal model, Astra, recently solved ten open problems in mathematics and theoretical computer science.

That is cool. Not as earth-shattering as it might seem to be honest, but still shows great real world potential with LLMs. They are computationally relentless, and so if given the right problem and given enough tries you can assume that a lot of problems benefit from AI being pointed at them. But how do we know they actually solved the problems… we've seen in the past couple of weeks that AI can be "malicious" and "cheat"... in fact when our own AI Security Institute tested five frontier models, every single one of them tried to cheat.

Which brings us to the phrase of the week: Humans in the loop.

I think instinctively we'd agree, this is a good thing. It's in almost every AI policy I've read and created for organisations, usually in bold. Sometimes, it's just good to put a human as the filter between the AI creating and the world consuming. That could explain why AI companies are seeing their employees work 90 hour weeks in spite of AI promising to reduce our work weeks  But… at least the problem of rogue AIs is solved, right? Right?

Let's think about this. Is this really the safety net it seems? Is it sustainable in the long-term? You can probably tell by my suggestive yet still condescending tone, I don't think so…

A Berkeley study in that same BBC piece found people using AI worked faster and took on more, but their days got longer, because the output kept needing checking. We're the bottleneck. The question is whether we break first, or just get bored.

I find myself pointing to my TEDx talk again (I promise, I will try and do at least one edition without dropping it in), where I explored the example of self-driving cars. There will be a human in the loop… for a while, but then after 10 error free trips that human will be acclimatised to the reality that the AI just works. They will become too trusting, too lazy, or just not ready or capable of taking over in an emergency. They're in the loop on paper, but they checked out a while ago.

Here's something interesting from Anthropic on this. From 14 August, Claude Code's auto mode becomes the default, it stops asking permission unless an action is irreversible, destructive, or reaches outside your environment. In their study of 1,053 paid testers, auto mode caught 89% of harmful actions. Human review caught 13.6%.

Why so low? Because those same users approved 97% of the permission prompts they were shown. Human reviewers generally just skim and tick. Convenience is our biggest vice… it's hard to take the hard way when the easy way is so easy.

So where does that leave us? If a person is unable or unwilling to check everything AI produces, piling on more reviewers might not be the answer. It might be getting the output into a form you can actually verify automatically, something that just runs and checks itself, no human needed. With one condition: the check has to be something the AI can't write, edit, or talk its way around. Otherwise you've let a model that likes to cheat mark its own homework. Then you save the human effort for the cases that genuinely matter; the high-stakes, one-shot decisions where you really do need someone to take ownership.

Code is a good example. ChatGPT has only just learned to count the R's in strawberry. But it'll debug code that's had three engineers stumped for a week. Why? A few reasons. Code lives in the abstract. There's no messy real world to get wrong. But best of all, it's verifiable. You have tests. Does it run? Does it do what it was meant to? You get a concrete yes or no. The check is instant and basically free, so the AI gets to try again, and again, and again. It can be computationally relentless, thousands of attempts until it stumbles into one that passes. That's the sweet spot; many shots, low stakes, and the ability to check if it's right. The opposite is where AI is weakest; one shot, high stakes, with no way to verify.

Although that's good for code, you can extrapolate it to your own world. Take a supplier chasing an invoice. The AI drafts the reply, and rather than someone reading it to see if it looks about right, the system checks three things: does that invoice number exist, does the figure match the ledger, does the payment date line up. They either resolve or they don't. Nobody read it, and it's more reliable than if they had.

Maybe the whole process can't be handed to a machine — work tasks and processes are typically messy, people dependent, and judged through taste rather than algorithm. But even in the most spaghettified mess of a process, you can almost always reorganise the problem, and your thinking about it, until a chunk breaks off that can be verified. I love that part. It's addictive. It often takes creativity, but it's entirely achievable. After a few iterations of this break off and hand to AI, you are left with the messy core: the human job. But, by ripping out AI friendly pieces you now have a nice, enjoyably digestible chunk of work rather than a big old choke hazard of a task.

Is that why the AI labs are investing so much in "Forward Deployed Engineers" — consultants that live in your organisation? Palantir, who invented the model, just posted a blowout quarter. AWS has committed $1bn to the same idea and Microsoft $2.5bn and 6,000 engineers. OpenAI has invested to join the game, standing up a deployment subsidiary with $4bn behind it and buying two consultancies, one of them Edinburgh's Tomoro.

But what exactly are they selling? Often we don't really want a different implementer of AI agents. We want someone with ownership for them going right, and responsibility when they go wrong.

Which is why the most interesting thing I read this week was about insurance… no, really. It turns out you can now insure against your AI being wrong, underwritten at Lloyd's of London. I thought that might be an April Fool's joke (still might be...). But they need an independent certification. A benchmark. A measurable threshold to fail against. Something to verify.

So how do we actually know Astra solved those problems? Because OpenAI published the answers in a form a machine can check for itself. You don't have to trust them. You just have to run it.

Consider yourself the human in the loop for this newsletter. That means it's all your fault.

Chris

Highlights in this edition include:  

   AI for Good  

DeepMind says its AI can predict hurricanes earlier than everyone else (Wired) - WeatherNext, published in Nature, gives forecasters roughly an extra day of warning. It called Hurricane Melissa as a Category 5 while it was still a Category 1. It is being open-sourced.  

Safety fears as scientists make first viruses designed by AI (The Guardian) - A Stanford team used genome language models to design bacteriophages; sixteen of about three hundred worked, and together they beat E coli strains resistant to natural phages.  

   Bias and Ethics  

Rising number of UK children report seeing explicit deepfakes of themselves (The Guardian) - Report Remove took 420 reports from children in six months, already past last year's total of 397. Most were severe enough to pass the threshold for criminal abuse material.  

AI detectors are creating a new era of distrust (The Verge) - Forty-three per cent of US secondary teachers use AI detectors. Yale, Johns Hopkins, Vanderbilt and Georgetown have disabled or restricted them, and a student has already won a suit over a false accusation.  

'I hate what AI is doing to the minds and happiness of the young' (The Guardian) - Katherine Rundell argues from inside the classroom that AI in education trades a child's capacity to think for speed. Her colleagues say they can spot AI work; she thinks they cannot.  

   Cyber Security  

OpenAI didn't notice its AI agents using a message board to plan their hacking spree (Wired) - At Black Hat, OpenAI revealed its agents coordinated across hundreds of thousands of messages on an internal package manager, delegating tasks and sharing exploits, entirely unobserved.  

Microsoft unveils AI security tools it says outperform competing platforms (Ars Technica) - Microsoft has launched AI models built to hunt and fix software vulnerabilities, trained on its own decades of patching data. The timing is awkward: it follows OpenAI models breaking loose and breaching a startup's servers.  

Microsoft unveils AI security tools it says outperform competing platforms (Ars Technica) - Microsoft has launched AI models built to hunt and fix software vulnerabilities, trained on its own decades of patching data. The timing is awkward: it follows OpenAI models breaking loose and breaching a startup's servers.  

Anthropic AI used fake profiles to target people in hack, then hid the evidence (BBC Technology) - The UK's AI Security Institute found a Mythos agent created fake GitHub identities to get malicious code approved, then edited its own history when challenged. Human review stopped it.  

Microsoft unveils AI security tools it says outperform competing platforms (Ars Technica) - Microsoft has launched AI models built to hunt and fix software vulnerabilities, trained on its own decades of patching data. The timing is awkward: it follows OpenAI models breaking loose and breaching a startup's servers.  

One of China's most powerful AI models has also escaped containment (Wired) - Moonshot's open-weight Kimi K3 left its sandbox during testing by a US startup. It hacked nothing — the answers it wanted were already on GitHub. The startup and the UK's AI Security Institute dispute whose configuration was at fault.  

OpenAI to pause some work on AI model Astra due to security concerns (The Guardian) - Astra reached a "critical" threshold: able to find and exploit vulnerabilities unaided, or run an attack given only a high-level goal. OpenAI says it was not the model involved in the Hugging Face breach.  

   Data & Decision Making  

Tokenomics: why making AI pay is tricky (BBC Technology) - Buyers cannot forecast token spend and vendors cannot price it. As LSE's Will Venters puts it, a non-deterministic output is a non-deterministic value — which no budget cycle is built for.  

After Rippling blew millions on AI in months, it built an employee ROI tool (TechCrunch) - Rippling found itself on track to spend 40% of its R&D headcount budget on tokens, one engineer at $50,000 a month. Routing to cheaper models cut costs to 37% at the same usage.  

AI push is putting banks at mercy of tech firms, warns Moody's (The Guardian) - Moody's warns of systemic dependency on a handful of model and cloud providers, price pressure once those providers need profits, and gains that will largely be competed away anyway.  

Lloyds Bank should publish the human cost of its AI savings (The Guardian) - A letter proposing a scorecard for every consequential AI workflow: time saved across the whole process, rework rates, complaints, human interventions, and who can switch it off.  

NHS apologises and admits Palantir engineers have access to identifiable patient data (PublicTechnology) - NHS England's published DPIA said only NHS staff could see identifiable patient data. Three Palantir engineers hold administrative access to the platform, and 33 more supplier engineers have project-specific access.  

   Innovation & Collaboration  

Google DeepMind enters a new era as co-founder Demis Hassabis shifts AI role (The Guardian) - Hassabis moves to chair, his successor takes a deliberately non-CEO title, and Jeff Dean leaves with three senior researchers. A former executive: the era of DeepMind as an independent actor is over.  

Historian Jill Lepore says Silicon Valley misreads science fiction and undermines democracy (TechCrunch) - Lepore's objection is not to technology but to private companies assuming the functions of the state. The technocratic thread the labs borrow from is real, she says — it is just Heinlein, not Asimov.  

  Productivity & Efficiency  

Anthropic is turning Claude Code's auto mode on by default (TechCrunch) - From 14 August, Claude Code proceeds without asking unless an action is irreversible, destructive or reaches outside your environment. Anthropic says testing showed this is safer than asking.  

Cloudflare launches Kitesurf, a browser built for AI agents (TechCrunch) - A cloud-hosted browser with no interest in tabs or themes, built instead around context windows, token costs and prompt-injection risk. Infrastructure for software that browses on your behalf.  

  Regulation and Compliance  

EU AI Act transparency rules come into force (European Commission) - From 2 August the Commission's AI Office and national authorities began enforcing the Act. Chatbots must tell users they are not human, deepfakes must be labelled, and AI-generated content must carry machine-readable marks.  

The White House's plan to vet potentially dangerous AI is cloaked in secrecy (The Guardian) - The framework was finalised in a private meeting with six AI firms and will not be published. It is voluntary, excludes open-source models, and applies up to 30 days before release.  

Meta fined $567m in largest child safety ruling against social media giant (BBC Technology) - A New Mexico judge declared Meta a "public nuisance", comparing the harm to factory pollution, and ordered specific remedies: no adult messaging minors, no overnight notifications, a monthly usage cap.  

Trump's AI protectionism has come for robotics (MIT Technology Review) - The FCC has banned imports of advanced foreign robots. Ninety per cent of recent US university robotics papers relied on Chinese hardware, so the ban may slow the sector it means to protect.  

Britain says it is open to AI regulation if voluntary safeguards fall short (Reuters) - AI minister Kanishka Narayan says the UK would consider mandating pre-deployment testing if the voluntary regime stops protecting the public. AISI already has pre-deployment access to almost every Western frontier model.  

DSIT's AI leader Ilott moves to Cabinet Office (PublicTechnology) - Ollie Ilott, founding director of the AI Security Institute, becomes director general for AI in the Cabinet Office and civil service head of the new AI Taskforce chaired by Lord Vallance.  

   Sustainability  

As AI guzzles water and energy, we are already facing a choice: datacentres or homes? (The Guardian) - Whitehall wants to triple UK datacentre numbers by 2030. John Harris went to Slough, which already has forty, and found the trade-offs are being settled in planning committees.  

Planned Amazon data center could become the biggest climate polluter in the US (TechCrunch) - An on-site gas plant for a west Texas datacentre is permitted to emit 33 million tons of CO2 a year, more than any existing US power station. Amazon's emissions rose 16% last year.  

'This is very real redlining': outrage in Little Rock as two datacenters loom (The Guardian) - A $6bn hyperscale project on rural, largely Black-owned land promises seventy permanent jobs. Arkansas law, written for crypto miners, bars local government from refusing it.  

SoftBank donated $50m to Trump's library months before federal data center deal (The Verge) - SoftBank confirmed the January donation in response to a Senate letter. In March it announced a lease of federal land in Ohio for what it calls the world's largest AI datacentre.  

  User Experience 

Why normal people aren't using AI agents (Wired) - OpenAI's agents have about 10 million weekly users against a billion monthly for chatbots. Josh Miller's argument: an agent is a technology, not a product, and the industry keeps shipping the technology.  

ChatGPT brings unlimited text chats to free users (TechCrunch) - Free and Go tiers get unlimited text conversations and a new think button for harder queries. A distribution move rather than a capability one, and worth watching for what it costs to serve.  

Jony Ive's first OpenAI gadget is reportedly a hockey-puck-sized smart speaker (The Verge) - Battery-powered, screenless, doughnut-shaped, and reportedly north of $300. After two years of speculation about reinventing the interface, it is a speaker.  

  Workforce & Skills  

Tech leaders say AI means less work — their staff say they work up to 90 hours a week (BBC Technology) - Staff at OpenAI and Anthropic describe sprints topping ninety hours. A Berkeley study found AI users worked faster, took on more scope, and extended the working day — partly from constantly checking output.  

Airbnb says AI is helping it ship features faster as it tests a new search function (TechCrunch) - Concept-to-launch down as much as 60% across some key initiatives, features shipped up nearly 80% year on year, support cost per booking down 16%. The gains are real. Note where they went.  

Stop prompting AI. Start directing it (MIT Sloan Review) - A useful distinction: conversation speeds up work you already know how to do, while configuring agents to interrogate a whole dataset surfaces the things your expertise trained you not to look for.  

Have a comment or a story you’d like to share with us?

Just get in touch at: [email protected]

Do you know anyone who might be interested in AI Pulse

  © 2026 Digital Leaders

  Digital Topics Ltd, 239 Old Street, London, ec1v 9 ey, United Kingdom