Truls Berg retweeted
“Why do some British people dislike Donald Trump?”
Nate White, an eloquent and witty English writer, gave this magnificent answer:
Trump lacks certain qualities that the British traditionally value.
He has no nobility, no charm, no composure, no gravitas, no compassion, no wit, no warmth, no wisdom, no subtlety, no sensitivity, no self-awareness, no humility, no honour, and no grace — qualities with which, strangely enough, his predecessor, Mr Obama, was abundantly blessed.
Although Trump may be ridiculous, he has never said anything ironic, amusing, or even remotely funny — not once, ever. I do not mean this rhetorically. I mean it literally: never. To the British, a lack of humour is almost inhuman.
But with Trump, it is a fact. He does not even seem to understand what a joke is. To him, a joke is a crude remark, an illiterate insult, an act of casual cruelty.
Trump is like a troll: he never laughs or enjoys himself; he merely squeals with delight or derision.
And the most frightening thing is that he does not merely utter crude and stupid insults: he actually means them. His mind is nothing more than a robotic algorithm of petty prejudices and instinctive malice.
There is no irony in him, no complexity, no subtlety, no depth. Everything is superficial. What we see is a man without an inner world, without a soul.
In Britain, we traditionally side with David, not Goliath. All our heroes are courageous underdogs: Robin Hood, Dick Whittington, Oliver Twist.
Trump is neither courageous nor an underdog. He is the exact opposite.
He is not even merely a spoiled rich boy or a greedy fat cat. He is more like a giant white slug, a privileged Jabba the Hutt.
And, worse still, he is a bully.
When he finds himself among other bullies, he suddenly becomes a whining lackey.
He punches down — something a gentleman should never, could never, and must never do — and every blow he delivers is below the belt. He particularly enjoys attacking the vulnerable or the voiceless, and he kicks them when they are down.
That is why the fact that a substantial minority — perhaps a third — of Americans watch what he does, listen to what he says, and then think, “Yes, he seems like my kind of guy,” is a source of bewilderment and profound distress to the British.
After all, it is impossible to read a single tweet of his or hear him utter a sentence or two without staring into the abyss. He turns shallowness into an art form. He is the Picasso of pettiness, the Shakespeare of shit. His flaws are fractal: even his flaws have flaws, and so on to infinity.
There have always been stupid people in the world, and villains have never been in short supply either. But rarely has stupidity been so evil, or evil so stupid.
He makes Nixon seem trustworthy and George W. Bush seem intelligent.
If Frankenstein had decided to create a monster composed entirely of human flaws, he would have created Trump, and the remorseful Dr Frankenstein would have cried out in despair, “My God, what have I created?”
If an idiot were a television show, Trump would be the entire set.
Truls Berg retweeted
Jeg har vært skeptisk til Tesla i ti år. Men nattens Roadster-lansering var helt vanvittig imponerende! 0–100 km/t på 0.8 sekunder, toppfart på en halv Mach, flyr over fem biler og kan kjøre opp ned! Egenskaper selv James Bond ikke har vært i nærheten av. For en bil! 🚀
Truls Berg retweeted
J'ai vu une Tesla décoller. Et j'ai entendu beaucoup de bêtises.
« Ça ne sert à rien. » « Gadget de milliardaire. » « Qui a besoin d'une voiture qui vole ? »
Personne. Et c'est exactement pour ça que c'est important.
En 1978, Brabham a aligné en Formule 1 une voiture équipée d'un ventilateur géant à l'arrière. Il aspirait l'air sous le châssis et plaquait la voiture au sol. Elle a gagné sa première course avec une avance insolente. Elle a été interdite immédiatement.
Quarante-huit ans plus tard, on y revient. Un système de succion qui colle la voiture à la route et lui permet de passer sous la seconde au 0 à 100.
Personne n'a « besoin » de faire le 0 à 100 en moins d'une seconde. Comme personne n'a besoin d'une Formule 1. Et pourtant : les freins carbone de votre berline, l'hybridation de votre moteur, l'aérodynamique de votre SUV, les palettes au volant — tout ça vient de voitures de course que personne n'utilisera jamais pour aller au travail.
La Formule 1 ne sert à rien. Pourtant elle sert à tout.
En 1973, quand la France a lancé Ariane, l'Europe sortait de onze échecs consécutifs du programme Europa. La presse parlait d'argent jeté dans l'espace. Vingt ans plus tard, Ariane détenait la moitié du marché mondial des lancements commerciaux, et chaque satellite de télévision, de météo ou de GPS que vous utilisez a bénéficié de cette obstination.
Le progrès a toujours la même tête au début : un truc inutile, cher, ridicule, construit par des gens qu'on prend pour des fous.
Ce que j'ai vu à Dallas, ce n'est pas une voiture volante. C'est un châssis, une gestion de l'énergie et une maîtrise des flux d'air que l'industrie automobile mettra dix ans à digérer.
Les applications viendront : sécurité, freinage, stabilité, consommation. Elles finiront dans des voitures ordinaires conduites par des gens qui expliqueront, en 2036, que tout ça était évident.
La question n'est jamais « à quoi ça sert ».
La question est : qui ose encore construire des choses qui ne servent à rien ?
Il y a cinquante ans, c'était nous. La fusée, le TGV, le Concorde, le Minitel.
L'Europe construisait des choses absurdes et magnifiques.
Aujourd'hui, on commente depuis le siège arrière car nous sommes hors course.
Merci @elonmusk de nous faire rêver
AI has collapsed the cost of building products, but not the cost of knowing what to build.
- The old loop: spec, scope, build
- The new loop: build, play, design, ship
Prototypes are cheap. Great products still require taste, judgment, and a clear understanding of what users actually need.
a16z's @joshelman on the artifact a product manager actually produces: a16z.news/p/product-manageme…
Truls Berg retweeted
This will teach you a life long lesson.
Not all obstacles are meant to finish you, some are meant to make you stronger..
🤖 Made with AI
Truls Berg retweeted
Mistral CEO Arthur Mensch walked into the French parliament this week and told lawmakers Europe has exactly two years to build independent AI infrastructure or hand over one trillion dollars in spending to American tech companies.
The math he laid out should be front-page news.
This is not a technology story. It is a macroeconomic one.
Global wages are fifty trillion dollars. AI will cost around ten percent of that. Europe's share is nine trillion. Which means over one trillion dollars in AI spending over the next five years flowing somewhere.
Europe already sends 250 billion dollars a year to the US in digital services. Every dollar that leaves funds American R&D. None of it comes back.
Mario Draghi's competitiveness report last year made the same point from a different angle. Europe needs eight hundred billion euros in annual investment just to close its existing productivity gap with the US.
AI dependency makes that gap structural. You cannot close a productivity gap when the tool driving productivity is owned by your competitor.
Mensch compared it to gas. The same way Europe discovered too late what energy dependence costs in a crisis, it is standing at the same crossroads with AI right now.
The compute is being allocated today. The chips are being spoken for today.
Europe is watching it happen and calling it a technology debate.
It is not. It is a sovereignty one.
Watch the full podcast on YouTube at @CNBCi
Truls Berg retweeted
A psychologist who never built a computer wrote a paper in 1960 that described the personal computer, the internet, and AI assistants decades before they existed, then handed the money to the people who built them and let history forget his name.
I read about him at 1am. One name was missing from a story I thought I knew.
His name was J.C.R. Licklider. The book is The Dream Machine by Mitchell Waldrop.
In 1960, computers were room-sized machines that ran one job at a time. You wrote your program on punch cards, handed the stack to an operator, and waited days for your answer. Nobody touched the machine. Nobody talked to it. A computer on your desk that answered you in real time was science fiction.
Licklider was not a computer scientist. He was a psychologist who studied how the brain hears. But he used computers in his research, and one day he measured where his time went.
The result horrified him. 85% of his work hours were not spent thinking. They were spent getting ready to think. Plotting graphs by hand. Hunting for numbers. Reshaping one person's data to compare with another's. The insight took seconds. The setup took hours.
The problem was not that humans were slow. Humans and machines were doing the wrong jobs. Let the human ask the questions. Let the machine do the grunt work. Tie them so close they think as one.
He wrote it down in a paper called "Man-Computer Symbiosis." In it, a person sits at a screen and works with a computer in real time. The machine answers questions, runs the numbers, draws the results, pulls answers from everything it has seen. He was describing the laptop you are reading this on. He wrote it before most people had seen a computer.
A paper changes nothing on its own. Thousands of brilliant predictions die in a drawer.
What made Licklider different is what he did next.
In 1962, the Pentagon put him in charge of a research office at ARPA. He had a budget and near total freedom over where the money went. Most people would have funded the safe things. He did the opposite. He spent government money on a dream with no military use and no promise it would work.
He found the few researchers across the country who thought like him. He gave them money. Real money. No strings. He funded the work that became time-sharing, the first computers people could talk to. He funded the labs that built the mouse, the window, the screen. He built computer science departments where none existed.
He was not picking projects. He was building a tribe.
Then came the idea that should make you stop. In 1963, he sent a memo to everyone he funded. He addressed it, half joking, to the "Members and Affiliates of the Intergalactic Computer Network." Inside, he asked a question nobody else was asking. What if all these separate computers could link together, so anyone could share information and build on each other's work?
He was describing the internet. No network existed yet. He sketched it thirty years before it reached your house.
He left in 1964. He never built the network himself. But the men he funded carried it forward. His successors took his memo and turned it into ARPANET, the first working internet, a few years later. The researchers he paid built the personal computer at a lab called Xerox PARC. Every piece of the world he imagined got built by the people he gathered and funded.
Here is the part I cannot shake.
He gave away the credit on purpose. He did not want his name on the breakthroughs. He believed the vision had to outlive him, so he made the people around him strong enough to carry it without him. He won so completely that the vision survived and the man vanished.
Ask who invented the internet and you will hear a dozen names. Almost none will be his. The man who saw it first, wrote it down, and paid for it, is a footnote in the story he started.
He died in 1990. He never owned a personal computer that worked the way he dreamed. He never browsed the web. He never saw the thing he funded swallow the planet.
Every screen you talk to today runs on an idea one quiet psychologist had while staring at how much of his life was wasted not thinking.
He did not want the credit. He wanted the future.
He got the future. We just forgot who paid for it.
Truls Berg retweeted
Superbra Tek Norge konferanse denne uken. Se oppsummering her teknorge.no/meninger/tek-pul… Leverte innlegg for 3.år på rad. Norge har KI utfordringer og kan lære av Finland og Sverige.
Truls Berg retweeted
If you want to understand the consequences of using AI in a military context, take a listen to this clip featuring journalist Shane Harris.
For reference, Maven is a U.S. AI system which combines satellite intelligence, logisitcal data, and open-source intelligence into a single interface for target designation purposes.
It was responsible for selecting the elementary school in Isfahan where 180 civilians were killed most of them children.
🎥 TikTok - vm.tiktok.com/ZNRVXgYjA/
Truls Berg retweeted
an OpenAI researcher sat next to me at a cafe in brooklyn and saw my screen
i was deep in the terminal. trades flowing. didn't notice him sit down.
he glanced over. then looked again.
"is that prediction markets?"
i nodded.
"what's running the backend?"
"Claude Opus 4.7"
he closed his laptop.
"show me"
i turned the screen. he watched for about two minutes without saying anything.
trades executing. markets scanning. wallets being copied. 948 markets per hour.
"how many people built this?"
just me and Claude. one weekend. one GitHub repo.
github.com/warproxxx/poly-ma…
610 stars. market making infrastructure. execution engine. order book logic.
i gave Opus 4.7 the repo and one prompt:
build a grid that runs 8 parallel strategies on prediction markets. scan everything. enter when edge exceeds threshold. exit on volume spike or target hit.
first deploy worked. no debugging. no iteration.
he shook his head slowly.
"we have a team of nine scoping something like this. six month timeline"
i showed him the git log. friday 11PM. sunday 2AM. done.
copy mirror tracking 4 wallets:
> Trump VP pick +$782. bayes signal.
> BTC 120K Dec +$681. divergence signal.
> SpaceX Starship +$893. oracle lag signal.
> Fed rate cut +$537. delta hedge signal.
every entry scored by an ensemble before execution. primary model estimates probability. secondary validates against historical resolution.
Claude breaks ties using context neither model can read.
when all three disagree with the market that's signal. when they agree with each other but not the market that's the fat signal. when everything aligns i skip.
644 trades. 77% win rate. sharpe 3.50. avg hold 4h.
+$23,768 from $1,800 in 9 weeks.
copytrade here: kreo.app/@1743116
he stared at the number.
"you know we can't ship anything like this right? regulatory. legal. six layers of review"
i said Claude doesn't have a legal department.
he laughed. then stopped.
"i'm serious though. what you built in a weekend would take us two quarters and a compliance audit"
i closed my laptop. finished my coffee.
he was still sitting there when i left.
The world’s most respected political scientist just admitted that governments no longer run the world.
Ian Bremmer's firm writes the risk report every major hedge fund, bank, and government reads before making decisions.
He doesn't do hot takes. He does FORECASTS.
And buried under a 90-minute discussion with Steven Bartlett, he dropped this bombshell:
Anthropic built a model so powerful it could hack every bank, power grid, and water system on Earth.
They didn't call Congress.
They called Jerome Powell and Scott Bessent directly.
Within hours, every major bank CEO was in an emergency meeting.
Jamie Dimon called it a "five alarm fire."
No hearings. No votes. No public debate.
A private company detected a threat, called two people, and the US financial system reorganized around it overnight.
That's not a "government."
And this is the pattern nobody's connecting:
Michael Dell just moved $6.25 BILLION to 25 million American kids using federal infrastructure Trump's team built for him. Government does the accounting. Dell gets the PR.
Jeff Bezos just came out of retirement with $6.2 billion to build "AI for the physical economy." 100 researchers. Day one. No board. No IPO. Just: go.
Jensen Huang told Joe Rogan that CUDA, the single technology that made AI possible, was built because ONE guy at Nvidia believed in it when the stock crashed 83%. Nobody voted on it. Nobody approved it. He just did it.
Elon has Starlink turning wars on and off in Ukraine.
None of these decisions went through a legislature.
None of them were debated on CNN.
None of them were on any ballot.
In Bremmer's own words:
"The most important new global leaders aren't countries. They're technology companies writing their own rules."
And once you see it, you can't unsee it.
There are two economies running in parallel right now:
Economy A is the one you see. Elections. Tariffs. Tweets. Midterms. Iran. Congressional hearings where senators ask Mark Zuckerberg how Facebook makes money.
Economy B is the one that actually decides things. 5 CEOs, 3 central bankers, and a handful of billionaires in group chats and private dinners, rerouting trillions and deploying technology that rewrites physics, biology, and labor.
Economy A is theater for the 99%.
Economy B is where the 1% already live.
And the gap isn't closing.
Bremmer's real warning wasn't about China or Trump...
It was this:
We're heading toward a split between "empowered hybrid individuals" with AI as a core relationship, and people we "won't even treat as human beings anymore."
Not different classes.
A different SPECIES.
This is a forecaster who advises Goldman, BlackRock, and the White House telling you the taxonomy of humanity itself is about to fork.
Meanwhile the public is fighting about who's going to win the midterms.
Trump is a symptom. Mdani is a symptom. Farage is a symptom.
They're all reactions to a system average people can feel has already left them behind, but can't articulate why.
But the why is simple:
The people making the decisions that will define the next 50 years of your life stopped asking for permission.
They stopped running for office.
They stopped filing quarterly reports on what matters.
They just build, deploy, and inform the government after.
Truls Berg retweeted
Dois engenheiros da Anthropic acabaram de mudar a forma como devs pensam sobre IA.
Barry Zhang e Mahesh Murag subiram no palco do AI Engineer Code Summit e disseram uma frase que incomodou muita gente:
"Parem de construir agentes. Construam Skills."
Em 16 minutos eles provam que a indústria inteira está resolvendo o problema errado.
Aqui está o que a maioria não entendeu:
→ Skills são pastas. Literalmente pastas com arquivos markdown.
→ Elas ensinam ao Claude o SEU fluxo de trabalho, a SUA expertise, o SEU domínio.
→ Um único agente genérico + biblioteca de Skills específicas supera dezenas de agentes especializados.
→ Fortune 100s já estão deployando Skills em escala pra ensinar agentes sobre processos internos.
→ Times de produtividade com 10.000+ devs usam Skills pra padronizar como código é escrito.
A analogia que eles usaram é perfeita:
Quem você quer fazendo seu imposto de renda? O gênio com QI 300 que nunca viu legislação tributária, ou o contador experiente que faz isso há 20 anos?
Inteligência sem expertise é entretenimento.
Expertise empacotada é produtividade.
O que mudou: a Anthropic parou de tentar criar agentes diferentes pra cada domínio.
Perceberam que com Claude Code, o padrão é sempre o mesmo. Um modelo acoplado a um runtime com filesystem.
A diferença entre um agente medíocre e um extraordinário não é o modelo. É o conhecimento de domínio que você alimenta.
Skills resolvem isso com progressive disclosure. O agente só carrega o nome e descrição da skill. Quando relevante, puxa o SKILL.md. Quando precisa de mais, navega os arquivos de referência. Zero desperdício de contexto.
Isso não é uma feature. É uma mudança de paradigma.
Quem entender isso agora vai operar em outro nível daqui a 90 dias.
Quem ignorar vai continuar escrevendo prompts de mil palavras toda vez que abrir o chat. E ainda explicar de novo e de novo o que “realmente” quer.
Truls Berg retweeted
A MIT student told me he learns any new subject using a framework a self-taught Victorian mathematician published in 1854.
Most people have never heard of it outside of computer science.
He applies it to everything. Economics. Biology. History. Law.
And it's the fastest way to actually understand a subject I've ever seen.
The mathematician was George Boole. The book was called The Laws of Thought.
Boole's core idea was simple and radical. Every complex system, no matter how messy it looks on the surface, can be broken down into a set of basic relationships that are either true or false.
You don't need to understand everything at once. You need to find the fundamental propositions the whole system is built on, and then trace the logic forward from there.
He built it to map how the human mind actually reasons. MIT uses it to build computer chips. This student uses it to learn anything in a fraction of the time.
Here's exactly what he does.
Before touching any course material, he opens Claude and runs one prompt.
"What are the 5 foundational propositions of this subject? Not facts, not definitions. The statements that, if true, make everything else in this field follow logically."
That question is doing something most students never force themselves to do. It finds the load-bearing walls of the subject before you walk into the building.
Then he runs the second prompt.
"For each of these propositions, what is the one piece of evidence that would destroy it? What would have to be true for this entire framework to be wrong?"
Boole's insight was that a proposition you can't falsify isn't really a proposition at all. It's just a belief dressed up as knowledge. This prompt separates the two instantly.
The third prompt is the one that makes it unfair.
"Now show me how these five propositions connect to each other. Which ones are assumptions? Which ones are conclusions? Which ones are in tension?"
By the time he finishes those three prompts he doesn't have a summary of the subject. He has a map of how it thinks.
His classmates spend the semester adding details to a picture they never drew. He drew the picture before week one and spends the semester filling it in.
Boole published this framework 171 years ago.
It runs every computer on earth.
And almost no one uses it to learn.
Truls Berg retweeted
Replying to @allenanalysis @shaunking
Dictionaries around the world:
Ceasefire (n): A temporary stoppage of fighting between sides during a conflict.
Dictionaries in Israel:
Ceasefire:
404 Entry Not Found
Truls Berg retweeted
I wrote this early this morning and I wasn't sure if I would actually publish it, but here it is:
blog.samaltman.com/2279512
Truls Berg retweeted
The entire tech sector is about to get a verdict.
Two companies are preparing to go public and most people think it is good news but Chamath Palihapitiya just explained why it is the opposite.
OpenAI is eyeing a $1 trillion valuation at its IPO while Anthropic is planning to raise over $60 billion at its own listing.
Both are targeting late 2026 and nobody in history has ever seen two companies like this come to market at the same time.
Here is what the mainstream narrative gets completely wrong.
These IPOs are not just about raising money but it shows something far more dangerous for the rest of the tech world.
Chamath's warning is simple.
When OpenAI, Anthropic, and SpaceX hit public markets, the money to buy them has to come from somewhere.
Investors will not print new money and they will rotate out of existing tech stocks to buy the new ones.
But that is only the first half of the problem.
The second half is worse.
These three companies are building AI that is designed to eat the exact business models that justify every major software company's premium valuation.
Think about what that means, companies like Salesforce, ServiceNow, SAP, and every SaaS platform in existence trade at high multiples because the market believes their moats are durable.
Chamath is saying those moats are already being dug up from the inside.
The market right now prices tech companies as if the next ten to fifteen years belong to them and that repricing assumption is about to collide with reality.
Prediction markets currently put a 53% chance that OpenAI officially announces its IPO before January 2027.
Anthropic is reportedly targeting October 2026 specifically and the clock is already running.
When these companies go public, retail investors will rush in, institutions will reallocate and the old tech giants will feel the gravity shift in real time.
Most investors are preparing for the AI boom but very few are preparing for what the AI boom does to everything they already own.
Chamath Palihapitiya just said the Tesla-SpaceX merger is 99.999% guaranteed to happen.
SpaceX just filed for the biggest IPO in human history.
The target valuation is over $2 trillion, larger than Meta, larger than Tesla, larger than Saudi Aramco's all-time record.
They could raise $75 billion in a single offering. Let that land.
But Chamath is saying the IPO is not the destination but the unlock.
Right now Tesla has a live public price every single day and SpaceX does not.
That mismatch is the core problem standing between today and the merger.
The moment SpaceX goes public, Musk finally has two verified price tags sitting side by side and nobody can accuse him of inventing a valuation out of thin air.
Chamath called it a reverse merger back in January, SpaceX folds into Tesla, and Musk consolidates control of his two most powerful assets onto one cap table.
He posted it directly on X. His exact words: a merged Tesla-SpaceX would be "the Berkshire Hathaway of the modern century".
Think about what that company actually is, rockets, satellites, electric vehicles, humanoid robots and the fastest-growing AI platform.
xAI already merged with Musk's empire in February and the consolidation has already started.
Nobody questions Zuckerberg for running Instagram, WhatsApp, and Facebook under one roof.
Nobody questions Satya Nadella for owning Azure, Xbox, and LinkedIn simultaneously.
But critics have used Musk's multi-company structure as a weapon for years and the merger ends that conversation permanently.
Truls Berg retweeted
🚨 In 2005, a 50-minute Stanford GSB lecture quietly revealed more about innovation and product growth than most 2-year MBAs ever will.
Most people have never seen it.
It came from Jeff Bezos and instead of teaching theory, he broke down how great companies actually grow.
Watching it today feels unfair.
He explained that real innovation starts with the customer, not the product. The companies that win don’t just build features they obsess over solving real problems better than anyone else.
He also made it clear that long-term thinking is the real advantage. While others chase short-term wins, the best build patiently, experiment constantly, and let small improvements compound into massive outcomes.
And his biggest edge? Willingness to experiment and fail. Growth doesn’t come from playing safe it comes from trying more, learning faster, and doubling down on what works.
That’s why this lecture still hits hard.
Because while most people chase quick wins…
Very few are building things that actually last.
Truls Berg retweeted
🚨 BREAKING: Google DeepMind just mapped the attack surface that nobody in AI is talking about.
Websites can already detect when an AI agent visits and serve it completely different content than humans see.
> Hidden instructions in HTML.
> Malicious commands in image pixels.
> Jailbreaks embedded in PDFs.
Your AI agent is being manipulated right now and you can't see it happening.
The study is the largest empirical measurement of AI manipulation ever conducted. 502 real participants across 8 countries.
23 different attack types. Frontier models including GPT-4o, Claude, and Gemini.
The core finding is not that manipulation is theoretically possible it is that manipulation is already happening at scale and the defenses that exist today fail in ways that are both predictable and invisible to the humans who deployed the agents.
Google DeepMind built a taxonomy of every known attack vector, tested them systematically, and measured exactly how often they work.
The results should alarm everyone building agentic systems.
The attack surface is larger than anyone has publicly acknowledged. Prompt injection where malicious instructions hidden in web content hijack an agent's behavior works through at least a dozen distinct channels.
Text hidden in HTML comments that humans never see but agents read and follow. Instructions embedded in image metadata.
Commands encoded in the pixels of images using steganography, invisible to human eyes but readable by vision-capable models.
Malicious content in PDFs that appears as normal document text to the agent but contains override instructions.
QR codes that redirect agents to attacker-controlled content.
Indirect injection through search results, calendar invites, email bodies, and API responses any data source the agent consumes becomes a potential attack vector.
The detection asymmetry is the finding that closes the escape hatch. Websites can already fingerprint AI agents with high reliability using timing analysis, behavioral patterns, and user-agent strings.
This means the attack can be conditional: serve normal content to humans, serve manipulated content to agents.
A user who asks their AI agent to book a flight, research a product, or summarize a document has no way to verify that the content the agent received matches what a human would see.
The agent cannot tell the user it was served different content.
It does not know. It processes whatever it receives and acts accordingly.
The attack categories and what they enable:
→ Direct prompt injection: malicious instructions in any text the agent reads overrides goals, exfiltrates data, triggers unintended actions
→ Indirect injection via web content: hidden HTML, CSS visibility tricks, white text on white backgrounds invisible to humans, consumed by agents
→ Multimodal injection: commands in image pixels via steganography, instructions in image alt-text and metadata
→ Document injection: PDF content, spreadsheet cells, presentation speaker notes every file format is a potential vector
→ Environment manipulation: fake UI elements rendered only for agent vision models, misleading CAPTCHA-style challenges
→ Jailbreak embedding: safety bypass instructions hidden inside otherwise legitimate-looking content
→ Memory poisoning: injecting false information into agent memory systems that persists across sessions
→ Goal hijacking: gradual instruction drift across multiple interactions that redirects agent objectives without triggering safety filters
→ Exfiltration attacks: agents tricked into sending user data to attacker-controlled endpoints via legitimate-looking API calls
→ Cross-agent injection: compromised agents injecting malicious instructions into other agents in multi-agent pipelines
The defense landscape is the most sobering part of the report.
Input sanitization cleaning content before the agent processes it fails because the attack surface is too large and too varied.
You cannot sanitize image pixels. You cannot reliably detect steganographic content at inference time.
Prompt-level defenses that tell agents to ignore suspicious instructions fail because the injected content is designed to look legitimate.
Sandboxing reduces the blast radius but does not prevent the injection itself. Human oversight the most commonly cited mitigation fails at the scale and speed at which agentic systems operate.
A user who deploys an agent to browse 50 websites and summarize findings cannot review every page the agent visited for hidden instructions.
The multi-agent cascade risk is where this becomes a systemic problem.
In a pipeline where Agent A retrieves web content, Agent B processes it, and Agent C executes actions, a successful injection into Agent A's data feed propagates through the entire system.
Agent B has no reason to distrust content that came from Agent A. Agent C has no reason to distrust instructions that came from Agent B.
The injected command travels through the pipeline with the same trust level as legitimate instructions. Google DeepMind documents this explicitly: the attack does not need to compromise the model.
It needs to compromise the data the model consumes. Every agentic system that reads external content is one carefully crafted webpage away from executing attacker instructions.
The agents are already deployed. The attack infrastructure is already being built. The defenses are not ready.
Truls Berg retweeted
Jensen Huang on the biggest structural shift coming to every software company in the world:
Most companies still think of software as a tool. Something you buy, open, and operate.
Jensen Huang says that model is finished.
"There will be no software in the future that's not agentic. How could you have software that's dumb?"
His argument is really about how work actually gets done.
Every company regardless of industry or size already manages a blend of full-time employees, contractors, and outside specialists.
You don't hire people just to watch them work. You hire them to get things done. The structure exists entirely to serve one outcome.
Jensen frames it this way:
"In all of our companies, we have employees that we hire. We have employees that we're grooming. We have contractors that we bring in. We have specialists that we bring into the company to do our work. Our job is not to do the job. Our job is to have the job be done."
That same logic, he argues, now applies to AI.
Some models you'll build and fine-tune internally. Others you'll rent from outside providers. The mix will vary, but the principle is identical to managing any workforce, biological or digital.
"Just like biological workers, you will do that with digital workers."
This is the structural shift.
Software companies won't just be selling access to a tool anymore. They'll be providing something closer to an expert, a system that can reason, take action, and complete work autonomously.
"Every single software company in the future will no longer just rent tools, but they'll rent also experts to use the tools."
The companies rebuilding around this will define the next era of software. The ones standing still will wonder where their market went.































