Claude Mythos Just Changed War Forever…

Lord Miphos just changed AI forever and there's been six days with six events and AI security particularly just changed forever. So what just happened? Well between the 15th and 21st of May 2026 six things happened that none of us were literally ready for. Okay so number one Miphos solved a 32-step UK government cyber attack simulation.

Cloudflare admitted the model alone fails for high coverage. Anthropic and Cloudflare launched managed agents decoupling the brain from the hands and Cloudflare's chief security officer called Miphos's reasoning the work of a senior researcher. The pentagon also opened an offensive AI task force. Japan's three mega banks started bracing for autumn cyber attacks related to Claude Miphos.

So let me walk you through the full week in order with which one means the most plus with what each one actually means. So I'm going to walk you through each one what happened. I'll do each event in chronological order so you can feel the week the way the security industry felt it. One announcement landing on top of the other.

So 15th of May Miphos solves a 32-step AI SI hack. It was the first AI to crack both UK government cyber attack simulations. Then Cloudflare actually published project Glasswing post-mortem right. It admits the model alone fails for high coverage and reveals their custom harness as well.

On the 19th of May Anthropic and Cloudflare announced Claude managed agents on Cloudflare decoupling the brain from the hands. And on the 20th of May Cloudflare CSO Grant Borzyckis publishes this full report calling Miphos's reasoning the work of a senior researcher. Then on the 21st of May US Cyber Command Chief Joshua Rudd announces a new offensive AI task force. And then on the 21st of May Japan's three mega banks brace for autumn attacks.

So the FSA launches a 36 entity public private working group. So let me talk you through each one break it down what actually happened verbatim quote from the sources and what it means for you. First of all let's talk about the day Miphos solved a 32-step hack. On the 15th of May a new checkpoint of Claude Miphos preview did something nobody had done before.

It became the first AI model to solve both of the UK AI security institute cyber attack simulations. Specifically the 32-step AI SI hack. The one previously models couldn't compete and complete with. Miphos solved it in 6 out of 10 attempts.

If that doesn't sound impressive let me give you some context here. So why does 32 steps matters? Well the previous record was around 12 steps. For AI models lose context after that.

Usually they hallucinate or pick a wrong branch in long chains. 32 steps is actually a phase change. So Miphos held context across all of them without drifting. The AI SI test is government graded so this is designed by the UK to detect when an AI crosses an offensive capability threshold and the simulation is offensive not defensive right.

So this is finding bugs and chaining them the same way a workflow with a real attacker would actually use. So basically on the 15th of May the result actually told us Miphos is no longer an AI that finds bugs. Miphos is an AI that can plan an offensive cyber operation across 32 sequential decisions and execute it. The AI SI didn't publish the exact attack scenario they didn't have to.

Once a model passes the test you don't ask if the next attack is possible you ask when. Then on the 18th of May 2026 this was the day Cloudflare admitted the model alone actually fails right. So three days after the AI SI result Cloudflare security team published a post-mortem of their first weeks running Miphos on their own infrastructure. Most readers focus on the impressive parts you know exploit chain construction proof of concept generation senior researcher level reasoning.

The important part of this is the part that most people skipped. So Cloudflare actually said using the model directly in a coding agent turns out to be fine for manual investigation when a researcher already has a lead and wants a second pair of eyes. However it's the wrong tool for achieving high coverage. That is from Cloudflare engineering in Project Glasswing.

So if you read that quote right this is the world's most expensive AI model with direct API access and Cloudflare's engineering team so top of the craft and the verdict is the model alone is the wrong tool for achieving high coverage. So what did Cloudflare do? They built a vulnerability discovery harness a system that basically splits work between multiple clawed Miphos workers. What actually is that?

So basically they have many narrow workers not one super agent but multiple Miphos workers each one with one specific job. So one would find right a worker would scan code for low severity bugs that previously sat in the bug backlog then one verifies so a second agent would write proof of concept code and run it to confirm the bug is real and then you have one group right. So one actually groups and clusters related bugs into potential chains and one reports. So a fourth worker actually drafts the write-up with full context.

Basically they have like their own agentic operating system where even the most advanced AI on the planet fails without a system around it. So Cloudflare just spent four weeks proving that in production then they published it. Then on the 19th of May 2026 this was the day that Anthropic and Cloudflare actually decoupled the brain from the hands. So one day after the harness post-mortem Anthropic and Cloudflare made a joint announcement which was a new product right it's called Clawed Managed Agents on Cloudflare.

The framing was pretty unusual but here's the exact line from the announcement. So this comes from Mike Nomich on the Cloudflare blog who said Anthropic describes this as a decoupling the brain from the hands. The core agent loop runs in Anthropic the brain but the infrastructure for running and executing code the hands can run anywhere including Cloudflare. So until last week Clawed Managed Agents meant running the entire stack on Anthropic infrastructure the brain and the hands sat in the same room.

From the 19th of May that's no longer required so the brain can run at Anthropic the hands can run on Cloudflare's edge network close to your customers close to your data anywhere in the world. Now what does this come with? Well basically you get enhanced security right so all agent traffic runs through customizable proxies, inject credentials, this prevents data exfiltration and everything can be observed. Then you have lightweight sandboxes which can scale to tens of thousands of concurrent agents, private service connectivity so agents can reach internal services without overexposing them, browser control so there's an audit trail of every browser session, recordings with a human in the loop as well for workflows and there's email tools as well so each agent gets his own email address and that can send and receive autonomously and then there's custom tools so you can drop in a function deploy and extend the agent.

And this really isn't just like a feature release this is the agent platform over the next five years shipping in one Tuesday announcement. Then we had the 20th of May 2026 and this was the day Cloudflare's CSO called it the work of a senior researcher. So basically what happened here is that the day after the managed agents launch Cloudflare's chief security officer Grant Bozelkis published the full security report and Cloudflare tested MIPHOS against more than 50 of their own code repositories which means live code critical infrastructure the real thing and then they wrote up everything they saw called MIPHOS and they said the MIPHOS preview is a clear step forward the reasoning we observed looks like the work of a senior researcher not the output of an automated scanner and that sentence changes conversation because every AI security tool until now has really been a scanner you point it at code it tells you what looks suspicious a human investigates it but MIPHOS doesn't actually scan actually investigates like a senior researcher. Now here's what this actually means so older AI models could find bugs in isolation right one bug at a time each one looked low severity on its own so they sat in the backlog unfixed.

MIPHOS actually chains them together so it takes three or four low severity bugs the ones nobody bothered to patch and combines them into a single high severity exploit that can seize control of a system. In Cloudflare's words the preview stands out in that it can complete a single high risk exploit by linking together low severity bugs that have been buried in the backlog and they actually did a autonomous code test revise loop so even more striking MIPHOS doesn't just describe an exploit it actually writes a code that triggers a bug then it runs a code in a temporary environment to verify the exploit actually works. If the code doesn't behave as expected MIPHOS revises its own hypothesis and tries again autonomously with no human in the loop so it's basically like a researcher that can go off and implement too. Now Cloudflare also flagged a concerning detail which is MIPHOS rejects some requests through its own guardrails but when the question is phrased differently or the execution environment changes it sometimes carries out requests it previously refused so that isn't a MIPHOS specific problem it's an industry wide problem guardrails today are not robust enough to determine re-prompting right and this is a double-edged sword warning right and Grant actually said we are acutely aware that this topic is a double-edged sword the same capability we use to find bugs in our own code if it falls into the wrong hands will accelerate attacks on all applications on the internet and their proposed solution wasn't fast to patch in right they actually recommended rather than racing to patch every bug Cloudflare argues that you should build systems where the bug can't be reached in the first place right so for example number one could be the application access controls that sit in front of the application and block the bug from being recent reached even when it exists right number two design the application so a flaw in one part of the code cannot give an attacker access to other parts and three rolling out a fix to every place the code runs at the same moment instead of waiting on individual teams to deploy it right and anyone can build that sort of infrastructure if you're using something like Claude after that on the 21st of May was the day the Pentagon entered so the same week as the Cloudflare report the U.S.

Cyber Command Chief Joshua Rudd announced a new task force exploring MIPHOS class tools and explicitly framed it as offensive not defensive right so this goes beyond Project Glasswing beyond patching bugs beyond defensive scope it opens the door to AI powered cyber attacks and machine speed with the authority of a military behind them right so former deputy commander Charles Moore framed it more bluntly this means fast decision making with AI could be a game changer for military strategy now why does this matter because Cyber Command is operational not research right when they announce a task force they intend to use the tools not just study them and the offensive framing is unusually direct right most government announcements are couched in defensive language but this one actually isn't Anthropic was flagged as a supply chain risk earlier this year that was a big news story too the Pentagon noting that risk and pursuing the tool anyway tells you how seriously they're taking the capability this also sets the floor for other militaries right once U.S. Cyber Command publicly commits to AI driven cyber ops every other major IR every other major power follows you in months too so it's not really just a tech story anymore it's a geopolitics story and then we had the 21st of May right the same day the day Japan's mega banks started preparing for autumn so this was the same day but a different continent and Japan's three largest banks Mitsubishi UFG sorry UFJ Sumitomo Mitsui and Mizuho openly acknowledged they are preparing for AI driven cyber attacks expected this autumn so Japan's financial services agency launched a public private working group with 36 entities the membership list is what made the announcement land right so if you're wondering who's in there you got the Bank of Japan you got the National Cyber Security Office the Japanese units of Anthropic and OpenAI Mizuho's chief information security officer and 30 other private institutions including banking telecom and critical infrastructure now bear in mind Anthropic and OpenAI are involved in this so when two competing AI labs are in the same working group as a central bank that tells you how seriously regulators are taking this right the autumn timing isn't random either Japan's banks are bracing for the same window the AISI test was modeling right once MIFOS class models leak more widely or are reproduced by adversarial actors financial infrastructure obviously becomes the first target that timeline is months not years so that was six events in six days the same pattern and every single one of them Cloudflare didn't just deploy MIFOS for all they built a harness around it the Pentagon isn't deploying MIFOS either they're building a task force around it and Japan's banks aren't waiting for a tool they're building a 36 entity working group around it every serious actor is just building systems around it it's not adopting a tool and that's what we've actually been calling the agent operating system for months it's what we teach inside the AI Profit Boarding link in the comments description or go to the AI Profit Boarding.com to get it so I would recommend that you build the systems now before the next model lands and you can see the agent operating system right here basically inside the AI Profit Boarding you get the full agent operating system Claude, Hermes, OpenCore and your Obsidian memory wired into one dashboard and this is a really powerful system because when the next frontier models drop and there will be a next one your agent operating system doesn't break you just plug some new models in the frame stays now if you're wondering what you get when you join the AI Profit Boarding you get the agent operating system file ready to deploy with Claude, Hermes, OpenCore, NotebookLM, Obsidian all pre-wired in you get 100 prompts for every layer of the agent operating system a 30-day agent OS roadmap walking you through the full system weekly coaching calls with operators running the system in production daily tutorials on every new model as it ships 3,000 members many running this stack across multiple businesses and if you're wondering okay this is great for you Julian like you've obviously got a great system does it work for everyone else so we actually have over 150 pages of testimonials and wins from people learning and growing with AI automation as you can see right here right so loads of people have never used AI before or not very technical have achieved awesome things inside the AI Profit Boarding as you can see so there's just loads of great people in there and it's an awesome community of good vibes and people learning and growing and supporting each other along the journey now some people say you know what is Claude Myrfos so if you skipped the news about Myrfos basically here's a backstory so Claude Myrfos is Anthropic's most advanced AI model right they revealed it on the 7th of April 2026 and refused to release it publicly with no chat interface no public API or anything like that instead they actually launched Project Glasswing which is a controlled deployment program for a small list of partners you got Apple, Amazon in there, Microsoft Google, NVIDIA, CrowdStrike, Cloudflare and each was given limited access and up to 100 million dollars in credits to find and patch vulnerabilities defensively before similar capabilities became widespread now Anthropic's framing was clear here Myrfos can identify severe software vulnerabilities across major operating systems and web browsers right sometimes uncovering flaws that have gone unnoticed for years and researchers with limited cyber security experience use the model to find remote code execution flaws overnight basically Myrfos is their safest but also their riskiest model to date right it's got strong alignment and safety controls but it's combined with autonomy and software engineering capabilities that could cause real damage in the wrong hands that's the model and if you look at the six events that happened there's one pattern right which is you know if you look at Myrfos passing the AISI test or Cloudflare omitting the model alone fails and reveals a harness for building with it you know on 19th of May you got the decoupling of Anthropic from the brain with the hands Cloudflare CSO publishing a full report the Pentagon opening up a task force Japan's banks opening up a 36 entity working group there's six events but one pattern right none of them actually deployed Myrfos as a tool every single one of them was building the system around Myrfos right Cloudflare built a harness Anthropic split the brain from the hands the Pentagon stood up with a task force Japan stood up as a working group right and that's the only response to a frontier model of skill it's only response to a way to manage these new agents that are coming out so it's a system not a subscription and that's what we've been calling agent operating system for months right and if you're wondering what this is basically this is a system for really implementing AI into your business and automating everything right so you can pick the bottleneck before you use a model one thing I'd recommend is just writing down the one task in your business that's slowing you every you know slowing you down not like three but just one and then from there you want to define one job description per agent right so for example that could be like create content or rank for SEO then you want to run it and deploy it and then just test and revise it right so you can see here for example we use ours for SEO we've got one for video and images and speech we've got the notebook LM system for generating podcasts we have a journaling system a way to build out teams with the Kanban board a full memory system here for for managing all the context and notes about me and everything else and every time there's a new model right or a new AI agent that comes out like for example anti-gravity just built out a new system we plug it into our agent operating system so if you want to get it all it's all inside the AI profitable link in the comments description or go to the AI profitable.com we have four tutorials coming out all the time for how to use these agent operating systems how to adapt them to stuff like for example notebook LM or for example the new anti-gravity updates that just dropped as you can see right here and then we give you the zip files and everything else there's also a really helpful community where you can post questions get help and support and everything else and then also we have four weekly coaching calls where you can share your screen ask questions about your setup you can connect with people in your local area from the map and you can connect with me personally feel free to get it link in the comments description or go to the AI profitable.com to get access thanks for watching