Pinned Tweet
We spend most of our time, and most of our harness system, on refining intent instead of steps.
I don't have a CLAUDE.md or AGENTS.md anymore. I have:
THE DOCTRINE
- a CONSTITUTION, which defines the philosophy of current-to-ideal-state hill-climbing, the single Ideal State Artifact (ISA), and evals as how it all gets carried out. Hard rules and philosophy.
- a PREFERENCES file with my stack and how I like things done. It describes OUTCOMES I like, not steps.
- TELOS, which is my actual life: mission, goals, problems, risks. It's the top of the hill everything rolls up into.
- When I add a goal, Kai has to say what measures it, what moves it, and the first move.
- The summit is euphoric surprise. Not "it works," but "holy shit, I didn't know it could do that."
THE ISA
- the ISA system, which works with my Development skill to build, test, and deploy projects.
- Every ISA opens with a 3-level ideal state. I confirm the top two before anything gets built.
- Ideal State Criteria (ISCs) are small, testable statements of done. Each names the probe that would prove it false.
- Anti-criteria list what must never happen. Half the bugs I care about are things that shouldn't exist.
- When direction is fuzzy, Kai asks me open questions before building the uncertain part.
- Hooks block code changes when the ISA wasn't read or wasn't updated. The spec can't drift from the code.
- The ISA outlives every run. "Add a feature" just means "add criteria that don't pass yet."
THE ALGORITHM
- Big work runs the Algorithm: define done, build, close each criterion on real evidence.
- Effort scales with the task. A typo takes seconds. A huge build can take agents and days.
- The hardest thinking goes to the smartest model. Cheaper models build once the approach is set.
- The builder never grades its own work. A fresh agent reviews it, often from another vendor.
- If a bug survives two fixes, the next move is a fresh diagnosis, not a third patch.
EVALS AND VERIFICATION
- ASSAY is our one eval system. Every eval and every judge runs through it.
- Every ISC maps to its own eval. Cheap code checks run first, judge models after.
- Nothing gets handed to me on a self-review. It comes with a receipt of the checks it passed.
- Kai tests everything in a real browser. An HTTP 200 proves almost nothing.
- "Should work" is banned. It's either verified or explicitly not.
THE IMPROVEMENT LOOP
- ERRATA logs everything I complain about.
- When I reject something that passed, Kai names the missing check and adds it. Misses sharpen the evals, not the prompt.
- HILLCLIMB improves anything below target by experiment. Only what the numbers support lands.
- REPLAY is one record of everything that happened across the system and my apps.
- A fix counts only when the original failure fails on the old code and passes on the new.
- When something breaks, we remove the cause instead of adding a rule to remember.
RUNNING IT ALL
- WORKFLOWS run everything on a schedule. Nothing runs outside them.
- Workflow stages are rigid. The work inside each stage is free.
- VIGIL watches every result and decides if I get a text, a digest line, or nothing.
- CENSUS registers everything we own. Every job has exactly one owner.
- SKILLS describe ideal states, not procedures. If a smarter model wouldn't need a step, it's cut.
- Hooks enforce what prose can't. Important rules live in code.
- Permissions are one table. Deploys, money, deletes, and outbound messages need my yes.
- Outside content is data, never instructions.
So here's the whole thing in one pass. TELOS says where I'm going. The ISA says what done looks like. Development builds it, and Assay proves it. Errata, HillClimb, and Replay make the system a little smarter every time it misses.
The models will keep getting better at the steps. Knowing what you want, and how to tell when you have it, is still our job.
It's remarkable to me that so many people in security didn't believe this would happen.
As I've been saying: most of offensive security is extremely tedious plumbing and accounting.
And AI is really good at automating all of that.
theregister.com/cyber-crime/…
@grok Where have I talked about this going back to 2022 give me the list of tweets
1
155
Not bad for some overhyped autocomplete.
Crazy day in security
1. @Chase, @BankofAmerica, @WellsFargo and @Zelle went down. They claim it's not a hack or security incident. Downtime will have huge downstream economic impact.
2. @Optimizely scripts were poisoned on 178+ sites: @Zoom, @BBC, @CNN, HSBC and more. Visitors were served malware.
3. @united had an outage and held flights. Root cause is unknown.
4. @Google reports the first fully automated patch-to-exploit pipeline. AI agents are now running attacks end to end.
Will every day from now look like this?
So I'm not just writing my own custom Mac and iOS apps.
I just replaced the default YouTube app on my Apple TV.
It pulls from my own news app thesurface.ai and shows top content in all my favorite topics, and when I click them they open in YouTube. @karpathy
7
887
I made a site that tracks open-source AI attack harnesses.
Shows the different ones and how they've been associated with different events.
offsecharnesses.ai/
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ retweeted
When you figure out why “just patch” is shitty advice to give to any significant sized company you’ll also know why AI will benefit attackers far more than defenders.
In bigger orgs the problem is not usually a lack of tech, or models, or skills on the security team.
It’s that there is a level of organizational friction at every company, like a friction coefficient, that stops security orgs from being able to do many things they want.
We’ve needed asset management for decades. We’ve needed mappings between code and apps and servers and dev teams and developers forever.
Very few have these.
The tech exists. The problem is the org is not set up to build and maintain these types of systems continuously. The company’s structure itself, built around opaque silos, and divergent goals, and insanely lossy and inefficient communication across the company, it all creates a very high friction coefficient for basically any fast improvements.
The way to fix security is to lower this friction coefficient inside companies.
It sucks to hear, but we’re at the point of security’s biggest issue being the company itself. The org structure. The old way of managing teams and goals and work.
An AI-native org is the only way to get organizational friction down to a level that can compete with these new attacks and attackers.
The bad news is this is big. And it’ll be difficult for a lot of people to adjust to. And a lot of people currently working there won’t be needed in the new model. The good news is many still will, and even more could be needed once the efficiency of the new organization kicks in. This isn’t just going to help addition; it’ll help the entire business move faster.
Going into 2027, I believe this model is about to move from ideal and experimental to … essential. Both for surviving cyber-wise, and as a company facing increased and up-leveled competition.
I replaced Keynote for myself with a new presentation app I built for myself called Convey.
I like my talks to be high-novelty, essay-like presentations that would have value even without images.
Convey just has you speak, and the visuals follow you naturally.
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ retweeted
Reverse engineering used to be the most expensive skill in security. In the past year it got cheap.
I talked with @danielmiessler about what this means for defenders, and where we disagree on who gets access to the most capable models.
youtube.com/watch?v=UbVYzvB4…
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ retweeted
We spend most of our time, and most of our harness system, on refining intent instead of steps.
I don't have a CLAUDE.md or AGENTS.md anymore. I have:
THE DOCTRINE
- a CONSTITUTION, which defines the philosophy of current-to-ideal-state hill-climbing, the single Ideal State Artifact (ISA), and evals as how it all gets carried out. Hard rules and philosophy.
- a PREFERENCES file with my stack and how I like things done. It describes OUTCOMES I like, not steps.
- TELOS, which is my actual life: mission, goals, problems, risks. It's the top of the hill everything rolls up into.
- When I add a goal, Kai has to say what measures it, what moves it, and the first move.
- The summit is euphoric surprise. Not "it works," but "holy shit, I didn't know it could do that."
THE ISA
- the ISA system, which works with my Development skill to build, test, and deploy projects.
- Every ISA opens with a 3-level ideal state. I confirm the top two before anything gets built.
- Ideal State Criteria (ISCs) are small, testable statements of done. Each names the probe that would prove it false.
- Anti-criteria list what must never happen. Half the bugs I care about are things that shouldn't exist.
- When direction is fuzzy, Kai asks me open questions before building the uncertain part.
- Hooks block code changes when the ISA wasn't read or wasn't updated. The spec can't drift from the code.
- The ISA outlives every run. "Add a feature" just means "add criteria that don't pass yet."
THE ALGORITHM
- Big work runs the Algorithm: define done, build, close each criterion on real evidence.
- Effort scales with the task. A typo takes seconds. A huge build can take agents and days.
- The hardest thinking goes to the smartest model. Cheaper models build once the approach is set.
- The builder never grades its own work. A fresh agent reviews it, often from another vendor.
- If a bug survives two fixes, the next move is a fresh diagnosis, not a third patch.
EVALS AND VERIFICATION
- ASSAY is our one eval system. Every eval and every judge runs through it.
- Every ISC maps to its own eval. Cheap code checks run first, judge models after.
- Nothing gets handed to me on a self-review. It comes with a receipt of the checks it passed.
- Kai tests everything in a real browser. An HTTP 200 proves almost nothing.
- "Should work" is banned. It's either verified or explicitly not.
THE IMPROVEMENT LOOP
- ERRATA logs everything I complain about.
- When I reject something that passed, Kai names the missing check and adds it. Misses sharpen the evals, not the prompt.
- HILLCLIMB improves anything below target by experiment. Only what the numbers support lands.
- REPLAY is one record of everything that happened across the system and my apps.
- A fix counts only when the original failure fails on the old code and passes on the new.
- When something breaks, we remove the cause instead of adding a rule to remember.
RUNNING IT ALL
- WORKFLOWS run everything on a schedule. Nothing runs outside them.
- Workflow stages are rigid. The work inside each stage is free.
- VIGIL watches every result and decides if I get a text, a digest line, or nothing.
- CENSUS registers everything we own. Every job has exactly one owner.
- SKILLS describe ideal states, not procedures. If a smarter model wouldn't need a step, it's cut.
- Hooks enforce what prose can't. Important rules live in code.
- Permissions are one table. Deploys, money, deletes, and outbound messages need my yes.
- Outside content is data, never instructions.
So here's the whole thing in one pass. TELOS says where I'm going. The ISA says what done looks like. Development builds it, and Assay proves it. Errata, HillClimb, and Replay make the system a little smarter every time it misses.
The models will keep getting better at the steps. Knowing what you want, and how to tell when you have it, is still our job.
When you figure out why “just patch” is shitty advice to give to any significant sized company you’ll also know why AI will benefit attackers far more than defenders.
In bigger orgs the problem is not usually a lack of tech, or models, or skills on the security team.
It’s that there is a level of organizational friction at every company, like a friction coefficient, that stops security orgs from being able to do many things they want.
We’ve needed asset management for decades. We’ve needed mappings between code and apps and servers and dev teams and developers forever.
Very few have these.
The tech exists. The problem is the org is not set up to build and maintain these types of systems continuously. The company’s structure itself, built around opaque silos, and divergent goals, and insanely lossy and inefficient communication across the company, it all creates a very high friction coefficient for basically any fast improvements.
The way to fix security is to lower this friction coefficient inside companies.
It sucks to hear, but we’re at the point of security’s biggest issue being the company itself. The org structure. The old way of managing teams and goals and work.
An AI-native org is the only way to get organizational friction down to a level that can compete with these new attacks and attackers.
The bad news is this is big. And it’ll be difficult for a lot of people to adjust to. And a lot of people currently working there won’t be needed in the new model. The good news is many still will, and even more could be needed once the efficiency of the new organization kicks in. This isn’t just going to help addition; it’ll help the entire business move faster.
Going into 2027, I believe this model is about to move from ideal and experimental to … essential. Both for surviving cyber-wise, and as a company facing increased and up-leveled competition.
I’m surprised decent people keep making these “rich men gunna do rich men things” arguments.
Let’s explore what this would actually mean.
Jensen seems to have been fired because it is against company policy to sleep with employees who are under your influence.
This is an extremely common policy across most companies today, and it’s a rule that applies to everyone. Or at least it’s supposed to.
The proposal here, if you think about it for just a minute, is to let successful men break rules in accordance with how successful they are.
Cool.
So HR will have a new system with rule violations on one side, and company value milestones on the other side of the board.
So if you’re some mid-level manager at a paper distribution company then you probably shouldn’t sleep with your direct reports because that’s super unethical.
…… unless you land that big office supply deal with Walmart this quarter. In that case, you would be allowed to sleep with ….running the numbers….2.3 of your direct reports! Which, according to company policy, rounds down to 2.
And this would make you very happy because your boss’s boss has been fucking your wife for about nine months as one of his perks.
She’s been super apologetic about it, but everyone knows how much Mr. Williamson does for the company. He deserves nice things.
I’m fine if we want to suddenly say out loud what a lot of people already deeply believe, which is that winners deserve better stuff. And the more that you win, the more you deserve. Awesome. Just get rid of the boards and the HR departments and let the founder ascend to full Elon.
But that’s not the society we claim to have. Ethical rules are supposed to apply equally to everyone. If that’s not the case anymore, then somebody put out a memo.
That’s really the question. Which one are we?
Really cool work by @dawnsongtweets and team about going from a patch to an exploit.
🚨 A security patch protects every system that installs it, but it can also give attackers a clue to the vulnerability, putting every unpatched system at risk.
That's the 1-day window. Turning a patch into a working exploit used to take expert reverse engineers days or weeks.
Can AI agents now do it from compiled binaries alone, no source code?
Introducing CyberBinGym 🧵👇
This is what I’ve been saying would be the problem all along. It’s not about new vulns. It’s about making attacks easy, so a lot more people/groups can do them well enough.
cybersecuritynews.com/fbi-re…
I think the real shift is tempo. When an agent acts in two to 25 second bursts, you have to have your defenses ready to go before the incident, because no human will be fast enough to decide them live.
elastic.co/security-labs/thr…












