Keith Weaver

Here's an idea I've heard floated that I strongly disagree with. "Security controls need to provide a great end user experience." Meaning there is little to no friction or added complexity for users. My opinion is that this idea is a bunch of BS.

Entra App Proxy continues to be one of the biggest hidden gems of Entra P1 For over a decade, we've been able to stop exposing risky apps to the Internet by routing through agents with outbound connections to Azure I don't care what vendor you use, just get it off the Internet

Happy Thursday Graph PowerShell now has 138 character cmdlets 🥳

I love getting gaslit by Azure all the time... Automation accounts and Function apps only support PowerShell 7.4, even though 7.5 was released 1.5 years ago and 7.6 was released a few months ago "Please consider updating it soon." 😒 No Azure, you consider updating it soon...

I think many executives currently look at AI-generated software and think: "Wow. It's already 90% there." What they often underestimate is that the remaining 10% is not 10% of the work. A senior developer reviewing a 5,000-line AI-generated pull request often has to spend hours just understanding the architectural choices, hidden assumptions and how all the pieces fit together. At that point you're no longer "adding a few fixes". You're reverse-engineering a codebase that appeared out of nowhere in five minutes. And many senior developers absolutely hate that kind of work. Most don't want to become full-time reviewers of machine-generated spaghetti while spending their days writing specifications and documentation for an AI instead of building software themselves. AI is extremely good at creating the impression that we're "almost there". But "almost there" can still hide enormous amounts of engineering, maintenance and human responsibility underneath.

Historical moment for red and blue teamers 🛡️ Azure Active Directory Graph Activity logs are now available🥳

Are the people in your organization rewarded to ignore security?

Entra Hardening Tip #6: Kill Standing Access 🔐 Admin accounts are the ultimate target. If these accounts stay active 24/7, you're leaving a door open for attackers to disable policies, create identities + credentials and hide in your tenant. To harden Microsoft Entra, you need Zero Standing Access. ⏰ Just-In-Time (JIT) Access Stop using permanent role assignments. Use Privileged Identity Management (PIM) so roles are only active when they are actually being used. This limits the time an attacker has to exploit a compromised account. 👥 The Second Admin Rule Don't let a single user activate a high-risk role alone. Require a second admin to approve the request. This forces an attacker to compromise two accounts to get anywhere. This "two-key" system breaks the attack chain and creates logs that are much easier to monitor. Checklist: ✅ Use PIM: Move your admins from "Permanent" to "Eligible." ✅ Log Justifications: Every activation needs a business reason. ✅ Enforce Approvals: Require a second admin for high privileged roles. ✅ Audit Often: Use Entra ID Governance to ensure access is still necessary. If they can't get the time or the second approval, they can't get to your data.

We must learn that we have to start finally turning away from the wonderful visions of technical magnificence solving cybersecurity and toward relentlessly doing the duller stuff that we can actually do that will gradually make a difference.

@merill Looks like you might have a typo? Target: User Action = Register or join devices should be Target: All cloud apps.

The fix: Create a CA policy → Include: All users + All resources → Target: User Action = Register or join devices → Conditions: Authentication Flows 👉 Device code flow → Access control: Block access 🔹 Exclude the specific apps that have a legitimate reason to use 'Device code flow'. For more info see - learn.microsoft.com/en-us/entra/id… 2/3

Entra Hardening Tip #3: Block device code authentication flow Device code flow is a feature that allows users to sign into headless devices like Teams meeting rooms and CI/CD pipelines. The problem: Attackers are increasingly using this sign in flow to phish users by tricking them into clicking a link and signing in with device code flow. The result is the attacker gets a valid token of the compromised on the attacker's remote device. 1/3

Lovable, the AI app builder with millions of users, has a mass data breach affecting every project created before their patch in November 2025. Any free account can access other users' source code, database credentials, AI chat histories, and real customer data through five unauthenticated API calls. The bug was reported 48 days ago on HackerOne. It's still open. Here's the breakdown: > The vulnerability is Broken Object Level Authorization. Lovable's API verifies Firebase auth tokens but never checks whether the requesting user actually owns the project. Any authenticated user can query any project. > @weezerOsint created a free account today and accessed another user's full source tree, including an admin panel built for Connected Women in AI, a real Danish nonprofit. The project was last edited 10 days ago with 3,703 edits this year. This is active work. > The source code contained hardcoded Supabase credentials (SUPABASE_URL, SUPABASE_PUBLISHABLE_KEY, SUPABASE_SERVICE_ROLE_KEY). The developer queried the database and got back real names, real companies, real LinkedIn profiles. Speakers from Accenture Denmark and Copenhagen Business School. Not test data. > Affected endpoints include /projects/{id}/*, /git/files, /git/file, and /documents. All return 200 OK for pre-patch projects. > Every AI conversation is stored and accessible through the same bug. Developers discuss database schemas, paste error logs, share credentials, and walk through business logic with the AI. All of it is readable. > Lovable patched new projects but left existing ones exposed. A project created in April 2026 returns 403 Forbidden. The same developer's older project, same API, same endpoint, same free account, same session, returns 200 OK with the full source tree. > The first HackerOne report (#3583821) was filed March 3, 2026. Lovable triaged it, shipped ownership checks for new projects, and left every existing project wide open. 48 days later, nothing has changed. > Employees from Nvidia, Microsoft, Uber, and Spotify all have Lovable accounts. The exposure is not limited to hobby projects.

Was recently in a 20-person problem call. Someone used an LLM against one of the logs with possible sources of issue. It of course came up with a theory to justify the prompting. It was not that at all, but it would have delayed them for hours, chasing it without human guidance..

🚨‼️We have published a guide on mitigating the imminent AI security apocalypse Step #1: buy our tool Golden age of snake oil I stg. And the ‘tool’ is a prompt.

Crazy how @ZackKorman understood this and it was enough to make him the mortal enemy of multiple large companies.

(comic) Engineer on a Sales call

Alright, I've stayed away from the Mythos stuff for a little bit. Going to comment on that, but AI as a whole. First, this AI industry is absolutely insane. I feel like I'm back in the 90s/2000s with innovation, but it's not tempered or methodical - it's pure chaos. Everyday there is some AI-dude-bro (or gal) clawing for followers claiming end of cybersecurity, end of software engineering, or this breakthrough changes everything. We're seeing the "streamer" effect of video games now exploding in every industry that hasn't been in whatever industry, but is now a AI-expert thus an expert in anything AI touches because they can prompt. Largely it's not, but what it is doing is requiring us to understand what AI will do to virtually every industry in the future. I'm sitting here right now at a conference I'm presenting at, and I spoke with an individual which was like man... I'm just trying to get through this SAP implementation at my company, I don't even know where to start with AI at the moment. We are still in the extreme early stages of what AI can do, and I think that's really the exciting part - we are at the infancy stages of this. Most enterprise can't handle AI, as most companies couldn't handle agile workflow when it came out either, it took time, but eventually adopted. I won't dive deep into the scalability of releasing AI to the masses based on compute, power, or subsidies because these are real hurdles we need to solve. As you can see with Claude's spike in popularity is causing them to have to dumb the model down upwards of 65% just to stay afloat (Claude is absolutely awful right now for coding - beware). Mythos is cool, really cool - but it's not earth shattering as claimed. The potential here we are seeing a glimpse of what can actually happen though. The ability to do extremely complex tasks, with insane context windows, and high-end reasoning. But, what we saw from other current frontier models including open LLMs, they were able to find the same issues, but had to be specifically targeted towards those code sections because of context limitations and complex task reasoning which was drastically improved in Mythos. What does this mean? Basically. Nothing. It's a lot of marketing hype - but it does prove out that as these models become smarter, it will inevitably produce much better code, be able to work in mind blowing fashions that we haven't seen before - but it will all come down to cost. Right now Mythos is extremely expensive because of the compute needed, and we may solve that over time, but it's not there yet. The subsidies right now means AI is not ready. Scale is our biggest bottleneck right now and until that's solved, the industry will not move as fast as it could. What's particularly impressive is how the open models are starting to perform on par (or better) with the frontier models and become way more efficient without restrictions (turboquant) as an example. Our ability to use near parity models on our own hardware will only continue to get better which is a huge threat for these companies. I at first looked at Cursor's implementation of Kimi as they were falling behind because it wasn't "their own model". That wasn't accurate, its that the open models are performing substantially better than from 6 months ago, and will soon be leading the charge or close to it. What does this mean for cybersecurity? The industry is changing rapidly, and I absolutely freaking love it. We needed a swift kick in the ass in this industry that was largely stagnant for the past 10-15 years. What used to be a handful of incredibly talented security researchers that knew systems internals, savants at reverse engineering and reading through millions of lines of ASM is now being afforded to the masses, but still has a long way to go. The reason AI is so good at doing this stuff is because they paved the way, and will continue to do so in different ways. Not eliminated or removed, enhanced and better than ever. AI is single handedly the largest theft of plagiarism that has ever happened in human history. I just got a 10K check from Claude for ripping off my Metasploit book to train its model to be smarter actually :P I am all for things that make the world a safer place. Our goal in cybersecurity is to fix the world, make it less harmful when using technology - we should be adopting this. Note that it's going to come with a ton of fluff, hype, doomsday predictions, people that are now AI exports or coding experts but have never written a line of code themselves. That's all to be expected if you have ever been to an RSA conference. AI will product meaningful change in an industry that needed it. Cybersecurity is much more than bugs or defects, it's protecting against risk. AI is a new emerging risk, it's going to keep us insanely busy right now, and for the foreseeable future.

📢 You already know FOCI, BroCI, and all the OAuth2.0 flows? But do you already know the secret token providers of Entra ID? In my latest research post I explore how you can, hidden from the Defenders, request new access token. cloudbrothers.info/en/avoid-entra… #EntraID #DefenderXDR

sam altman watching ChatGPT hallucinate live on stage is the funniest thing i've seen all week the CEO of OpenAI, on stage, in front of everyone, watching his own AI just make things up in real time and his face says it all this is the guy telling us AGI is coming soon btw

Should an hourly Intune remediation run within the first hour after assignment? It sounds like it should...Right? And that is exactly why this remediation behavior keeps confusing people. Once you see what actually kicks off the first execution, the whole thing looks very different. I pulled that apart in our latest @PatchMyPC blog. patchmypc.com/blog/why-an-ho… Ow btw... Happy Easter Monday, everyone. #MSIntune #Intune #PowerShell #Windows11

You can have all the fancy security tools you want. If your IT/Security team don’t care, if they are just going through the motions, stuff will get by, you will have incidents because of the mistakes they make.

youtube.com/shorts/LNzxTvh… Oh I love this short. @ZackKorman