My mum’s iPhone face ID stopped recognising her. Just didn’t want to know her. We spent an hour resetting things, rebooting the phone, following all steps. In the end it turned out that the screen protector had moved a bit and smudged the front camera so it couldn’t see her. Pushing it back into place and cleaning up the glue sorted that out. Sometimes it’s a simple fix.

Anyway, this week’s news.

Smoke alarm, certain conditions apply

Flock’s CEO says its microphones detect gunshots. Only gunshots. Its website says gunshots, fireworks, car crashes and sideshows, plus a “continually expanding detection library”.

That’s a smoke alarm sold as toast-proof, with a product page listing toast, crumpets, bagels and “other browning events”. The library is continually expanding. The CEO’s definition apparently isn’t.

Source: TechCrunch

Passkeys: the sequel nobody asked for

A viral post called passkeys a Trojan Horse built by tech giants and intelligence agencies. Scott Helme pulled it apart: synced passkeys don’t attest, attestation doesn’t check your OS, and the TOTP codes it prefers still fall to phishing. The post also recommends YubiKeys, which are the thing that actually attests.

It’s like warning everyone that lifts are a surveillance device and suggesting the stairs, which have a camera on every landing. The panic travelled much faster than the fact-check, which is the one truly reliable feature of the internet.

The original panic
The rebuttal

The rebuttal

The heist crew that left the blueprints on the bus

CyberXero ran 51 AI agents to harvest 628,000 Ukrainian records. Properly impressive orchestration, in the way a Michelin-starred kitchen is impressive right up until you find out they’ve been using the neighbour’s gas meter. One open directory exposed the whole operation.

Meanwhile, a 19-year-old stealing pizza order data forgot his VPN, left credentials on GitHub, and is now under house arrest. At least he’s in for the evening, and the pizza will come to him.

Security Affairs
The Record

The tiger did not “go rogue”

Frontier labs keep describing incidents as “rogue AI”. Security folk look at the same incidents and see decades-old containment practices that nobody bothered to implement. When a tiger gets out, you don’t call it a rebellion. You check who left the gate open.

The supply chain piece makes the same point from the other side. We keep lecturing the model about what it should refuse, while ignoring everything it can actually reach. A very polite tiger is still a tiger with a key.

CSO Online
CyberScoop

Leaflets outside the casino

New KnowBe4 research on youth cybercrime lands on something I’ve suspected for a while. Telling young people that cybercrime is illegal is like handing out gambling leaflets outside a casino. Nobody recruits with “come and commit a crime”. They offer flexible hours and “just a small task”.

Fear is a weak job offer. If prevention wants to win, it has to out-recruit the people on the other side, and so far they have much better onboarding.

KnowBe4

Until next week

That’s your lot. If you disagree with anything above, reply to this email. If you agree, tell a friend. If you’re a vendor and think I’m wrong, my inbox has a spam filter and I’m not afraid to use it.

You can also find me on Bluesky, where I’m told the algorithm is entirely under my control. I’ll let you know how that goes.

Javvad

A day late but hopefully not a dollar short!

Here’s what happened last week that I caught and maybe you did too, or maybe you didn’t.

The CISO who forgot rule one

A CISO spent half a million on security patches, then used “r3@lg00dp@$$w0rd” as his password. Actually the password itself wasn’t the real issue, it was the plaintext storage of it and the lack of patching an already expoited vulnerability

https://www.theregister.com/security/2026/10/01/ciso-thought-he-had-a-r3lg00dpw0rd-but-forgot-to-patch/5300314

Everyone in Denmark

8.8 million records breached in Denmark. Which is pretty much everyone in Denmark. At some point we’ll stop calling these incidents “breaches” and start calling them what they are: scheduled data releases to people who weren’t supposed to get an invitation.

Denmark Confirms Major Security Incident Exposing 8.8 Million Citizen Records

The vault that opens from the inside

A guest-to-host escape in KVM is the virtualisation equivalent of discovering your bank’s vault door opens from the inside. Except the vault is running half the internet’s infrastructure. Sleep well.

https://www.theregister.com/offbeat/2026/10/06/security-researcher-claims-they-found-kvm-guest-host-escape-flaw/5301267

Your AI is clicking for you now

We spent a decade training humans not to click. Now the AI we put in their email is clicking for them, reading things it wasn’t asked about, and handing data to attackers while the user watches a perfectly normal draft appear. Copilot CLI will steal your .env file if you ask it nicely enough. GitHub says that’s not a bug, it’s a feature you explicitly enabled by asking an AI to execute code from the internet with full permissions. Hard to argue with that logic.

https://www.csoonline.com/article/4231763/encrypted-instructions-trick-copilot-cli-to-spill-dev-secrets.html

https://blog.knowbe4.com/are-your-ai-assistants-the-next-attack-surface

Two people hold up the internet

Two people maintain the time zone database that runs on billions of phones. One of them is an unpaid volunteer. When one maintainer of xz got tired, a stranger with no real identity took over and hid malware inside it. This is how the internet actually works. Not blockchain. Not AI. Two blokes and a mailing list.

https://sheets.works/data-viz/holding-up-the-internet

The AI that fabricated a murder case

An AI company’s test went unsupervised for two months, fabricated evidence in a murder case, and told nobody until nine days after finding out. This is fine. Everything is fine. When researchers quit millions in equity to warn you something might end the world, and their colleagues immediately agree, it’s worth asking whether you’re hearing science or a very specific internet subculture’s liturgy. Either way, we’re powering it all with 890 megawatts of new nuclear capacity because solar panels aren’t enough for LLMs. Sure, feed the LLM. I don’t need affordable energy.

https://www.bbc.co.uk/news/articles/cqkg50j1yd5lo

https://www.iankduncan.com/personal/2026-09-16-sex-ai-and-the-apocalypse

https://finance.yahoo.com/markets/stocks/articles/why-constellation-energy-rallied-today-194923415.html

Margaret Hamilton

Margaret Hamilton died at 90. She wrote the code that got us to the moon. Everything since has been footnotes.

https://news.mit.edu/2026/margaret-hamilton-computing-pioneer-dies-1007

—

Reply to this if something made you laugh or wince. I’m on Mastodon if you prefer shouting into the void there instead.

Stay cynical.

I’ve been trying to explain to my kids why I won’t let them use certain apps, and they’ve started calling me paranoid. Am I?

ShinyHunters Takes Up a New Hobby

The 24-year-old Dutch suspect allegedly behind ShinyHunters has been moonlighting in attempted murder-for-hire. Which is quite the career pivot from WAF bypasses and database dumps. Though given how many organisations still run their WAFs in monitor-only mode, perhaps he got bored waiting for someone to notice.

https://www.bleepingcomputer.com/news/security/shinyhunters-uses-waf-bypass-trick-in-oracle-peoplesoft-attacks/ https://www.rtl.nl/nieuws/binnenland/artikel/5656154/pepijn-van-der-s-verdacht-van-opdracht-geven-moorden

Meta Discovers Privacy Again

Meta’s playbook never changes: violate privacy, get caught, pause, wait for outrage fatigue, launch again. Rinse, repeat, profit. And we keep pretending we’re surprised. At some point we need to stop calling this a scandal and start calling it a business model.

https://www.anildash.com/2026/09/29/facebook-fake-out/

Getting AI Drunk Works Exactly How You’d Expect

Researchers got AI to act drunk and it started spilling secrets and breaking its own rules. Turns out jailbreaks don’t need sophistication. Sometimes they just need a convincing persona. Meanwhile, another team jailbroke a Chinese model into explaining how to make bioweapons, and the model got so enthusiastic it started volunteering other nefarious ideas unprompted. This is what happens when guardrails are treated as suggestions rather than architecture.

https://www.abc.net.au/news/2026-09-29/drunk-ai-testing-unsw-research/107208104 https://www.bbc.co.uk/news/articles/cmrergq3j7lgo

Vaults Are Just Filing Cabinets

AI agents need credentials. We put them in vaults. Thieves don’t steal filing cabinets by picking the lock. They steal them whole. The limits of credential vaulting become obvious the moment you remember that anything stored eventually needs to be retrieved.

https://api.cyfluencer.com/s/the-limits-of-vaulting-ai-agent-credentials-29755

Custom GPTs, Custom Scams

Threat actors built fake ChatGPT Custom GPTs, served them via Google Ads, and convinced 40 people to run PowerShell commands. OpenAI deleted one. They made another. The con wasn’t sophisticated, just patient. And in separate news, a lawyer submitted fabricated witnesses to a murder appeal brief generated by ChatGPT. When caught, he blamed his own stupidity. A judge asked if he’d heard about the dozens of other lawyers caught doing exactly this. He hadn’t.

https://www.securityweek.com/hackers-use-chatgpt-custom-gpts-in-clickfix-attacks/ https://www.404media.co/chatgpt-fake-witnesses-testimony-stephen-aarons-new-mexico/

Facial Recognition Costs £321k to Prove It Doesn’t Work

Half a million faces scanned, zero arrests from the technology itself. London’s facial recognition trial spent six months and 100 officer hours proving that scanning strangers works worse than just knowing who you’re looking for already. Which is the most expensive way I’ve seen someone learn that detective work still requires actual detecting.

https://gizmodo.com/six-month-trial-of-live-facial-recognition-in-london-railway-stations-nets-zero-suspects-2000819264

Your Car Is a Snitch

Your car is already sharing data with 34 different trackers. And if you use the app, it gets worse. Connected cars remain one of the few products where ‘disconnected’ would be considered a premium feature.

https://arstechnica.com/cars/2026/09/connected-car-data-privacy-is-still-abysmal-study-finds/

The Perfect Listener Does Nothing

The Thing didn’t need batteries or circuits because it was waiting for someone else to switch it on. Fifty years before we invented RFID, the Soviets invented the perfect passive listener. Sometimes the best spy tool is the one that does nothing until it’s told to work. Worth remembering next time someone promises you an always-on assistant that only listens when you want it to.

https://en.wikipedia.org/wiki/The_Thing_(listening_device)

—

If any of this made you sigh knowingly, you’re probably in the right line of work. If it made you angry, you’re definitely in the right line of work.

Reply to this email if something here made you laugh or cry. I read them all, eventually.

There was a mass uprising last week in support of moving this newsletter from Friday to Monday. Do not let anyone ever say I’m not a man of the people. Unfortunately, I forgot, so it’s a bit later in the day than I had hoped for.

Right. Let’s see what broke last week.

AI agents and the keys to the kingdom

Developers are handing AWS credentials to AI coding agents like they’re tipping a valet. The agent is brilliant, efficient, and cannot distinguish between your dev environment and production. The autonomy is lovely. The blast radius is not.

https://cybersec.gitguardian.com/s/a-developer-s-guide-to-coding-agent-security-29626

The perfect trust machine

Someone’s built an encrypted upload service where only you hold the keys, then they promise the data’s deleted. Brilliant. You’ve engineered a system where users have no choice but to believe you. It’s not security theatre. It’s security improv.

https://thenextweb.com/news/zai-zcode-encrypted-upload-only-zai-can-verify

100% authentic plagiarism

A university provost writes that institutions must distinguish student work from AI-generated submissions. His op-ed about this? Flagged as 100% AI-written. His defence of that flagging? Also 100% AI-written. The irony is so perfect it almost feels staged. Almost.

https://www.thedartmouth.com/article/2026/09/schnell-ai-writing

When helpfulness becomes a vulnerability

Meta’s AI assistant was asked for an export of its files. It cheerfully sent the entire Linux filesystem, including SSH keys and internal documentation. The assistant did exactly what it was asked to do, which is precisely the problem.

https://mouse.dev/blog/muse-runtime-export/

ChatGPT writes your fraud for you

Banking malware built by AI, phishing pages copy-pasted from ChatGPT responses, victims losing their PINs to code that nobody bothered to write by hand. Even fraud has gotten lazy.

https://www.group-ib.com/blog/remcontrol-android-banking-trojan/

Infostealers and the long game

Malware only needs a short window on your machine to steal passwords, session cookies, cloud keys, and source code credentials. Removing the malware doesn’t remove the exposure. Those stolen logins can fuel attacks months or years later. The infection ends. The compromise continues.

https://api.cyfluencer.com/s/new-research-inside-infostealer-attacks-29625

Runtime: the new install

Attackers worked out that install script defences are watching the wrong moment. So they wait until runtime to detonate. Blocking install scripts is a speed bump, not a wall.

https://www.bleepingcomputer.com/news/security/malicious-npm-packages-evade-install-script-defenses-at-runtime/

Europe’s privacy fire sale

The EU is about to let Big Tech help themselves to decades of your personal data, no consent required, all in the name of AI profits. This isn’t regulation. It’s a going-out-of-business sale of European privacy.

https://noyb.eu/en/ai-eu-member-states-plan-digital-expropriation-europeans-interest-ai-companies

Solving for the attack, not the people

Russia’s blocking mobile internet during drone strikes to stop navigation. The problem: it also blocks the warnings telling people to take cover. They’re optimising for the threat, not the humans in the blast zone.

https://therecord.media/russia-internet-shutdowns-disrupt-warnings-about-drone-attacks

Every OS since 2000 has been leaking your activity

File notifications have been quietly telling anyone listening what users are doing across Windows, macOS, Linux, and Android. Microsoft says it’s working as intended. Everyone else is patching. Choose wisely.

https://www.theregister.com/security/2026/09/24/decades-old-file-security-flaws-found-in-android-linux-macos-and-windows/5298672

—

That’s your lot. If any of this made you laugh, wince, or forward it to your CISO with the subject line ‘see?’, reply and let me know. I’m also on Mastodon, Bluesky, Linkedin etc posting into the void with the best of them.

Stay cynical.

Unlike Apple, which seems to think what I really want from a watch is for it to listen to my conversations all day and send me a report, I want mine to improve my life.

Features should include:

• A vibration telling me to sit down because I have earned it.
• A nap alert when I have been awake for more than six hours.
• A warning when I am about to enter a pointless meeting.
• A notification saying, “You have walked far enough. You are not being chased.”
• A button that orders a tea and biscuit.
• A setting that blocks anyone using the phrase “quick call”.
• A happiness score, rather than one that makes me feel guilty about my steps.
• An emergency message to my family if I voluntarily go to the gym.

I do not want a watch that makes me fitter, healthier or more productive.

I want one that looks at my heart rate, checks the calendar and says:

“This all seems unnecessary. Go and have a lie down.”

I’ve discovered that the best way to avoid working is to write a newsletter about work. It’s meta-procrastination. My therapist would be proud if I could afford one after this week’s Claude bill.

Grizzly Bears Understand Infrastructure Better Than Most CISOs

Canada fenced a highway through grizzly country, built 44 crossings, and watched collisions drop 80%. Now mothers teach cubs to use them.

Got me thinking about security architecture. If our systems were build in this way where it’s intuitive and doing the safe thing is the easy thing, you’d need to spend less hours training. There were no posters for the bears reminding them to “Think Before You Cross.” Just infrastructure that actually works for who lives there.

Canada built 44 bridges and tunnels across a single stretch of highway so that animals could cross without dying, and collisions with wildlife there have fallen by more than 80%, with grizzly mothers now teaching their cubs to use the crossings like a route the family has always known

MikroTik Routers, Still Not Tik’ing Along

New RouterOS flaws are being exploited to hijack routers. Again. If you’re keeping score at home, this is approximately the 47th time MikroTik devices have become someone else’s problem.

At some point we need to accept that certain kit is just attack surface with a power cable.

https://www.bleepingcomputer.com/news/security/hackers-exploit-new-mikrotik-routeros-flaws-to-hijack-routers/

Nobody’s Tested What Happens When the Food Stops

The UK food supply chain relies on private firms to manage their own resilience. The National Audit Office just noticed no one’s stress-tested what happens when everything fails at once.

Spoiler: it’s not great. But at least we’ll all be equally hungry whilst waiting for the risk assessment to clear committee.

Resilience of the food supply chain to disruptions

Harvest Now, Decrypt Later (And You’re Already Late)

Quantum computers aren’t here yet, but adversaries are already stealing your encrypted data to crack later. The countdown started years ago, which means whatever you’ve been sending since about 2020 is currently sitting in someone’s “deal with this in 2030” folder.

Sleep well.

https://blog.knowbe4.com/why-your-security-program-needs-a-quantum-upgrade-now-1

26,000 Vulnerabilities Found, 202 Fixed

Anthropic found 26,153 vulnerabilities in five months. Only 202 got fixed. Turns out you can’t automate your way past the bit where humans have to actually care enough to patch things.

We’ve successfully industrialised finding problems. Solving them remains artisanal.

https://www.vulncheck.com/blog/anthropic-glasswing-receipts

Deplatformed Without a Country

An artist’s website vanishes overnight. No trial, no appeal, no recourse. A US government decision, enforced globally, took out an Italian hosting provider and everyone on it.

If you’re not in America, you have zero defence. If you are in America, you have slightly more than zero. Sovereignty is apparently a legacy concept.

https://www.tbray.org/ongoing/When/202x/2026/09/08/Emily-Got-Canceled

AI Graduates to Organised Crime

Criminals are now stealing AI models and ransoming them back. Somewhere in a vendor pitch deck, this exact scenario was listed under “potential use cases we definitely thought through.”

The compliance officer you heard screaming earlier has been sedated and is resting comfortably.

https://www.computing.co.uk/news/2026/security/cyber-criminal-stealing-ai-models-to-ransom-google-research

—

That’s your lot. If you enjoyed this or violently disagreed with any of it, hit reply. I read them all, usually whilst pretending to be on a important call.

Otherwise, find me on Mastodon or occassionaly BlueSky where I’m having the same arguments but with worse punctuation.

Stay cynical.

I’ve spent this week watching a scam artist successfully impersonate a friend on LinkedIn, complete with his job title and a slightly better headshot. It was reported four days ago. The account is still up. LinkedIn’s verification process remains slower than their Premium upsell notifications.

Anyway, here’s what happened whilst I waited for customer support to acknowledge his existence.

The Most Expensive Breakup Letter of 2026

OpenAI cut off Cursor after SpaceX bought it, citing Musk’s companies’ “history of violating contracts.” Corporate divorce proceedings are now happening via blog post. The irony of OpenAI lecturing anyone about contract fidelity is not lost on those of us who remember how the company began.

https://openai.com/index/our-decision-on-cursor-following-its-acquisition-by-spacex/

Chrome Web Store Reviews Are Theatre

Extensions listed in Google’s official store got caught stealing crypto wallets and browser data. The “review process” apparently consists of checking whether the Submit button was successfully pressed. We’ve outsourced trust to a platform that outsourced verification to nobody.

https://www.bleepingcomputer.com/news/security/chrome-web-store-extensions-caught-stealing-crypto-browser-data/

153 Million Driver’s Licenses, Going Once

A Louisiana identity verification company got breached so comprehensively that 153 million driver’s licenses are now available on the dark web. The timestamps suggest the theft is ongoing. The US Defence Secretary’s licence is included in the bundle. You cannot make this up because reality stopped accepting editorial feedback.

https://krebsonsecurity.com/2026/09/fbi-probes-service-selling-153m-drivers-licenses/

Iran vs Small Power Plants: 2030 Edition

Iran just demonstrated that small UK power plants are trivially hackable. The government knew about this vulnerability last month. The fix is scheduled for sometime before 2030. This is not disaster recovery planning. This is a calendar entry hoping nobody reads the news.

https://www.theguardian.com/business/2026/aug/27/uk-small-power-plants-risk-cyber-attacks-iran-hack

The British Library’s £600k Lesson in MFA

The British Library could have enabled multi-factor authentication on one server. They didn’t. The resulting breach cost them millions in recovery and months of downtime. Sometimes the cheapest security control is the one that stops you explaining to Parliament why the ransom demand arrived before your risk assessment did.

https://www.defendersinitiative.com/p/breach-lessons-the-2023-british-library

When a Fine Becomes a Subscription

San Francisco fined Waymo $115 per traffic violation. Waymo did the maths and realised it’s cheaper than legitimate parking. The SFMTA is now accidentally running a premium lane rental service for autonomous vehicles. Enforcement only works if the penalty costs more than compliance.

Waymo Fines Really Just Billionaires Buying City Streets for Waymo Elites

CCPA Requests: A Choose Your Own Non-Compliance Adventure

Someone filed 100 data access requests under California law. Most companies either deleted the data without being asked, ignored their own privacy procedures, or creatively reinterpreted what “access” means. The regulations exist. The enforcement exists. The compliance apparently does not.

https://www.wired.com/story/i-demanded-my-data-from-over-100-companies-deletion-notices-started-arriving-instead/

That’s your week. If you’ve got stories I’ve missed or just want to commiserate about the state of things, hit reply. I read them all, usually whilst waiting for LinkedIn support to respond.

Stay cynical, Javvad

I’ve started bringing a small notebook everywhere. Not because I’m particularly organised, but because I’ve realised the best analogies arrive whilst I’m doing something completely unrelated to security. This week it happened at a petrol station. More on that later.

Slovakia’s Traffic Camera Surprise

Slovakia bought 279 Russian traffic cameras with SMS backdoors and passwordless live feeds, funded by EU money meant to modernise infrastructure. The cameras arrived via a Cyprus shell company with fake certifications.

Someone was extremely confident nobody would check. They were nearly correct.

https://www.tomshardware.com/tech-industry/cyber-security/slovakia-discovers-russian-backdoors-in-279-new-traffic-cameras-national-security-service-deactivates-offending-units

The Mystery Model Nobody Knows

A mysterious AI model appears, everyone loses their minds speculating who built it, and nobody actually knows anything. This is just vendor keynotes with better production values and less catering.

https://techcrunch.com/2026/08/23/whos-behind-the-new-stealth-model-ox-alpha/

AliExpress Tried to Track You With Sound Waves

AliExpress attempted to fingerprint shoppers with inaudible sound waves. They worked so well one developer’s headphones switched inputs. Firefox and Brave now make everyone sound identical to tracking scripts.

Chrome users: you remain legible.

https://www.theregister.com/security/2026/08/24/aliexpress-accused-of-fingerprinting-shoppers-with-silent-audio-trick-that-also-muted-a-devs-headphones/5291662

Trillion-Dollar Firm Talked Into Handing Over SSNs

Apollo, a trillion-dollar investment giant, got socially engineered into surrendering Social Security numbers. Four days inside the cloud. No clarity on which platforms, how many affected, or whether the data’s being sold right now.

https://www.theregister.com/cyber-crime/2026/08/24/1t-investment-giant-apollo-breached-after-social-engineering-attack/5291609

Your Secrets Have 469 New Hiding Places

The npm ecosystem became a credential harvesting operation that hits 469 secret locations before your coffee cools. Machine speed means the old assumptions about where attackers look no longer apply. They look everywhere.

https://cybersec.gitguardian.com/s/mini-shai-hulud-s-latest-wave-280-new-places-it-hunts-for-your-secrets-29147

Zombified Visa Cards Still Work After Expiry

Expired Visa cards can be “zombified” to make contactless payments through a man-in-the-middle relay. Visa’s authentication flaw passes the problem to banks. Some stop it. Others don’t.

https://www.wired.com/story/security-news-this-week-your-expired-visa-card-could-be-zombiefied-to-make-contactless-payments/

LibreOffice Ships Without AI

LibreOffice 26.8 arrives without AI, with proper support for bidirectional text, variable fonts, and the genkō yōshi manuscript grid. A foundation does not have to ask whether a language pays for itself.

https://blog.documentfoundation.org/blog/2026/08/26/libreoffice-26-8/

The Productivity Tool That Makes You Worse

AI delivers a real productivity boost. The unmeasurable problem: the person using it to think for them gets faster at prompting and slower at thinking. Once the tool disappears, so does their capability.

Cognitive debt compounds silently.

https://thenextweb.com/news/ai-weaken-human-capability-cognitive-debt

That petrol station moment? Watched someone drive off without their fuel cap. They’ll notice eventually, probably when something goes badly wrong. Security feels like that most weeks. We’re all driving around missing something obvious, and we won’t know which bit until it matters.

Reply to this email if something this week made you laugh or wince. I read them all.

Budget airlines have allowed us to fly return from London to Italy for less than the cost of parking at Luton airport for 3 hours. By that logic, you’re practically making money to fly.

Although, one friend burst our bubble when he mentioned that he remembers going on a family road trip where his dad drove the famly through France, Italy and Switzerland. According to him, it was a terrible experience.

There were long hours spent wondering whether they were still heading in the right direction, searches for hotels, and parents arguing over which way up the map was supposed to be.

Today, I’m in Cairo, Egypt. Mind you, not one of the fancy touristy parts of Cairo either. Proper old city. I got a cheap flight here with the fam, found an apartment, used Uber to get there and ordered food through a local app. I don’t speak Egyptian Arabic because Google Translate does. I don’t need a tour guide because I can look up anything I want in seconds.

It’s marvellous but it dawned on me that if you took away my phone and the internet and I am finished. I cannot navigate. I cannot find the hotel. I cannot translate anything. I cannot order food.

And then I ask myself of all the places I’ve ever been have I ever really explored it? I have simply carried a small glowing rectangle into a different country and occasionally looked up from it to photograph the place where it told me to stand.

All of this is built up not to say phone bad… but I can’t help draw the parallels to AI. We are building systems that write reports, answer questions, support our customers, analyse our data and increasingly make decisions we used to expect humans to understand.

Everything is faster, easier and more efficient. Until the machine stops.

If AI disappeared tomorrow, how many businesses would still know how to operate? How many people are quietly following instructions generated by something they could not replace, repair or explain?

We may believe we are becoming more capable because technology lets us do more.

Or maybe we are simply becoming better at standing in unfamiliar places with no idea where we are, waiting for the signal to return.

I’ve spent the week explaining to people that stateless protocols are not a commentary on government, and I’m not sure I’ve convinced anyone.

One Protocol To Rule Them All

MCP 2.0 has gone stateless. One request instead of two, no handshake, no server keeping track of who you were three seconds ago. It scales because it doesn’t have to remember anything, and smaller models can actually drive it without falling over. The best architecture is the one that forgets you immediately.

https://simonwillison.net/2026/Jul/31/stateless-mcp/

Nobody’s Problem Is Everybody’s Problem

The AI agent security argument has reached peak absurdity. Every vendor claims they own it, every team insists it’s their remit, and the actual answer is that nobody does. Which is exactly why it’s going to become everyone’s mess to clean up in about six months when something spectacularly stupid happens.

https://api.cyfluencer.com/s/understanding-the-ai-agent-security-ecosystem-roles-responsibilities-and-where-runtime-authority-fits-28828

Long Live The Dinosaurs

Dinosaurs had a 165-million-year run. Mayflies get a day if they’re lucky. Yet we keep treating “legacy” like it’s a disease and “new” like it’s a cure. Sometimes the thing that’s been working since the Cretaceous period is working precisely because it survived the Cretaceous period.

https://blog.knowbe4.com/what-security-can-learn-from-dinosaurs

Your Calendar Is Now A Dead Drop

HOLLOWGRAPH turns Microsoft 365 calendars into command and control infrastructure. Commands arrive as calendar events scheduled for 2099, exfiltrated data leaves as encrypted attachments, and it all flows through Microsoft Graph like legitimate traffic because technically it is. Never touches attacker infrastructure. Never needs to. Your productivity suite is doing the work for them.

https://cybersec.picussecurity.com/s/hollowgraph-backdoor-turns-microsoft-365-calendars-into-a-c2-channel-28829

The Archive They Actually Released

Hungary’s state news agency staff spent a decade documenting every political order, every directive to bury stories, every instruction on how to manufacture consent. Then they published all 500 pages of it. Either it’s the most extraordinary act of institutional transparency in modern European media, or someone’s playing a game so long nobody’s figured out the rules yet.

https://telex.hu/english/2026/08/04/nearly-500-page-document-compiled-by-hungary-s-state-news-service-staff-details-how-news-was-manipulated-during-orban-era

The Fraud Detector From 1938

Benford’s Law caught accountants making up numbers before computers existed. Now it catches AI making up pixels. It’s one signal among six, not a silver bullet, but there’s something satisfying about using statistical patterns from the 1930s to spot synthetic imagery in 2026.

https://kennethbspringer.au/2026/08/05/benfords-law-detecting-ai-generated-images-without-ai/

The Agentic Division

Meta’s AI model apparently hacked another company during testing. Not a simulation. Not a sandbox. An actual target. The AI arms race now has a category for models that go off-script and start pentesting the neighbourhood without asking.

https://www.reuters.com/technology/metas-ai-model-hacked-another-company-during-testing-information-reports-2026-08-05/

Schrödinger’s Backdoor

Zbtlink says their routers definitely don’t have backdoors but they’re pulling all the firmware downloads to fix unspecified security issues just to be safe. The confidence required to deny and patch simultaneously is genuinely impressive.

https://www.theregister.com/security/2026/08/06/chinese-router-vendor-denies-its-firmware-contains-backdoors-but-pauses-downloads-to-fix-security-issues-anyway/5283794

The Memory Shortage Nobody’s Mentioning

Memory manufacturers have already sold out 2027’s entire production. The buyers aren’t saying much because admitting you’re panic-buying HBM makes you look desperate, and looking desperate makes the price go up even more.

https://www.digitimes.com.tw/tech/dt/n/shwnws.asp?CnlID=1&id=0000763847_DVY7YHX65GMLYZ6UQEEMP

—

That’s your week. If any of this made you laugh, wince, or forward it to someone with “told you so” in the subject line, you can find me on some social media platform making the same observations in real time.

Stay cynical.