←
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️29,904 posts
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ avatar

ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️

@DanielMiessler

Building Vector: The Enterprise AI Operating System @ ourvector.ai and LifeOS: The Human Magnification OS @ ourlifeos.ai | Ex: Apple, Robinhood, HP, Army

◷ Joined March 2007▧ 2,294 Photos and videos
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ avatar
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️@DanielMiessler·Nov 16

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.

lauren@potetoNov 15

interesting how well this post aged what are your predictions 1 year from now?

ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ avatar
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️@DanielMiessler·12h

I think I have a serious chance of getting a major prediction wrong in 2027. I've been thinking that AI would seriously start hitting security startups and VCs that fund them in 2026 and definitely in 2027. And I also anticipated all this .... whatever this is ... as soon as models and harnesses got good enough (especially open ones). Somehow I didn't put these together into 2027 potentially being a great year for VCs and cybersecurity companies, just because so much security will be needed by so many. I'd say my intuition is still mostly the same, that there will be massive consolidation, but now I see a possibility of the opposite. More hype, more panic, and more investment into more companies. What do you think will happen?

ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ avatar
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️@DanielMiessler·16h

I cannot express to you how bad this is given how dominant Anthropic has been lately, and how much we've been waiting to see OpenAI's response. I have close friends who work there, so I've held off on giving my analysis until now, but at this point it's not helping anyone to avoid commenting publicly. Here's what I see happening, as a comparison of OpenAI to Anthropic. Both companies are in constant chaos. That's just a function of how insane and awesome the work is that they're doing, combined with how fast everything is moving in AI, in tech, etc. So that's understandable, and not really anyone's fault. But they're handling it in very different ways. Anthropic has the same founding members, and the same board as when they started. They have a clear vision, articulated constantly by a clear leader. So while they have chaos, they handle it better because they filter new employees for those who think similarly to the company and its mission. And they're actual believers in the mission...which (again) is very clear to everyone. OpenAI does not have a clear mission that I can see, after watching since the end of 2022 (and a bit before). They have one on paper, and they sometimes articulate it well. And Sam and Greg are often good at articulating it in various settings. But it isn't being consistently broadcasted and pushed down in like a semi-religious sort of way the way that Dario does, or Jobs did, or Jensen did. Kind of a tangent here, but what I've learned over many years in tech and business is that this type of quasi-religious leadership is kind of the only way to keep a team/company/project going steady over long periods of time. Sam and Greg are brilliant, obviously, and they have a million great ideas, but that unfortunately manifests as constant change of focus, vibes, and direction. And I feel like the result is people not feeling grounded or psychologically safe. Like the focus could just change and your whole team could get moved and/or nuked for the next shiny thing. From what I can tell, and have heard, Anthropic has tons of this kind of chaos and churn as well, but it's just different when the MAIN thread hasn't changed even once. Same message. Same people. Same religion. Same mission. With that type of consistency it's possible to handle a lot more external chaos without people freaking out. So basically I'm getting really worried about OpenAI. They're raised way more money. They feel way more over-extended (although I'm not an expert on the details). And in general they just seem way more top-heavy and vulnerable. I think this perfectly explains all this going on with the model and harness wars as well. OpenAI can at any moment drop a banger and shock us all because they have so much money and so much talent, and they're trying so many things. But it's not focused. It's not stable. It's haphazard. Again, just my read. Meanwhile Anthropic, at least compared to OAI, is just slow and steady. More careful investments. Fewer bets, on fewer threads. And as a result they keep shipping super high-quality stuff. Going back to the leadership thing, there is just a massive advantage to having a true believer type in charge of everything, e.g., Elon, Jobs, Jensen, Dario. It provides an incredible force of direction to the whole company and everyone in it. I think Sam has vision too, but it feels more sterile to me. More tactical, or calculated. Still super smart, but it doesn't feel like quasi-insane conviction like you see with these other leaders. I mean it's not even necessarily a bad thing, since there are obvious downsides to being too religious as well. But I think this is hurting OpenAI right now because, given how crazy things are, it could be that this is the only type of leadership that can keep something this big from falling apart. Anyway, this is by no means a prediction. Langchain and N8N and Jev were all we could talk about for 9 minutes and now they're kind of off the board. And a few months ago Anthropic was the one on the ropes in the harness world. So things could change. But my current bet is on Anthropic being at least one of the winners in all of this, just because of their leadership and direction. And it would not surprise me if OpenAI over-rotated onto its face at some point in 2027. Like a total meltdown. It also wouldn't surprise me if they launched a chip, or a personal AI device that supplants the iPhone, or a new model, and instantly became the only thing everyone is talking about. Anyway, hope this gives someone some sanity if they've been thinking similarly. And I'd love to hear if/where my analysis is incorrect.

Post attachment
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ avatar
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️@DanielMiessler·16h

The developer pulled the repo on October 8, and as of early November there were forks, mirrors, a commercial ARTEX PRO, and 1,600+ exposed consoles. My guess is we see a lot more of these FLASHes. If you haven't inventoried your partner portals and old business-unit apps lately, I'd start there, because that's where these agents are finding the way in.

ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ avatar
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️@DanielMiessler·16h

One of the attack characteristics they ask you to report is "evidence of multiple sub-agents overwriting each other or generating duplicate requests." I love this one, because the agent bugs are the forensic tell, and anyone who's run a multi-agent harness knows exactly what that looks like in a log.

Post attachment
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ avatar
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️@DanielMiessler·16h

In the activity the FBI catalogued, it enumerated over 110,000 Active Directory objects and specifically listed which Windows machines didn't have CrowdStrike Falcon. But at least once, its attempts to move laterally to the domain controllers got blocked by endpoint protection. So the basics still work. Honestly kind of a relief.

Post attachment
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ avatar
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️@DanielMiessler·16h

It does best on the stuff nobody watches. Partner and supplier portals, mobile backends, employee apps, legacy services, third-party integrations. This is what I meant by AI only having to beat a 3 out of 10. A mediocre agent that never gets bored finds your forgotten partner portal before your team does.

Post attachment
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ avatar
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️@DanielMiessler·16h

Most attacker-run instances use DeepSeek Flash models. The report says why: good enough, cheap, and no visible guardrails once you tell it it's doing a pentest. ARTEX plugs into Claude and GPT too. The people actually running it went cheap.

ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ avatar
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️@DanielMiessler·16h

The FBI is calling these "super intelligence (SI) agents." In an official FLASH. I wrote last month that I was torn about renaming AI to super intelligence. Did not expect the Bureau to adopt it this fast.

Post attachment
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ avatar
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️@DanielMiessler·16h

The FBI re-issued this FLASH this morning and changed one claim. The Nov 13 version said the ConnectWise campaign "enabled access to more than 2,250 downstream tenants." Today's version says it "resulted in attempted sign-ins" to them. I'm glad they corrected it, because access and attempted sign-ins are really different things.

Post attachment
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ avatar
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️@DanielMiessler·16h

The FBI just put out a FLASH on ARTEX, the open-source agentic pentest framework attackers have been using since at least August. I read the whole thing, and a lot of it is stuff I've been worried about for a while. Some intrusions went from initial targeting to privileged access in under three hours. One campaign took personal data on about 67,000 people from South Korean financial firms. Another went through ConnectWise Cloud Backup customers and tried sign-ins at more than 2,250 downstream tenants.

Post attachment
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️ avatar
ᴅᴀɴɪᴇʟ ᴍɪᴇssʟᴇʀ 🛡️@DanielMiessler·23h

Don’t just tell me how good your harness is. Show me what it produces. - sites - mobile apps - designs - apps - tools - products - services - OUTCOMES Are you helping people with it? Is it making you money? Has it improved your life? Judge a harness by its output.