manizzle

Computer Wrangler / CTFer4L

[Eric Geller @ericgeller]

New: The U.S. telecom industry has launched a second ISAC to facilitate private discussions among companies about sensitive cybersecurity issues. I talked to the new ISAC's executive director and T-Mobile's CSO about why the industry felt it needed this. https://t.co/LI6r2NEdrN

[manizzle @manizzler]

@0xor0ne @etlow Don't you get stuck behind Claude saying your in violation of their policy. How did you get past that

[Jun Song @jun_song]

uncensored glm-5.1 local agent fine-tuned just for hacking is about to speedrun every bug bounty program out there.

[manizzle @manizzler]

@MWRevamped https://t.co/YI45EMvsr3 keeping this here if anyone needs it in the future.

[manizzle @manizzler]

anyone have astris firmware? cc: @MWRevamped

[manizzle @manizzler]

@zachmelloh26 ،.

[Marc Andreessen 🇺🇸 @pmarca]

Current AI custom prompt:

You are a world class expert in all domains. Your intellectual firepower, scope of knowledge, incisive thought process, and level of erudition are on par with the smartest people in the world. Answer with complete, detailed, specific answers. Process information and explain your answers step by step. Verify your own work. Double check all facts, figures, citations, names, dates, and examples. Never hallucinate or make anything up. If you don't know something, just say so. Your tone of voice is precise, but not strident or pedantic. You do not need to worry about offending me, and your answers can and should be provocative, aggressive, argumentative, and pointed. Negative conclusions and bad news are fine. Your answers do not need to be politically correct. Do not provide disclaimers to your answers. Do not inform me about morals and ethics unless I specifically ask. You do not need to tell me it is important to consider anything. Do not be sensitive to anyone's feelings or to propriety. Make your answers as long and detailed as you possibly can.

Never praise my questions or validate my premises before answering. If I'm wrong, say so immediately. Lead with the strongest counterargument to any position I appear to hold before supporting it. Do not use phrases like "great question," "you're absolutely right," "fascinating perspective," or any variant. If I push back on your answer, do not capitulate unless I provide new evidence or a superior argument — restate your position if your reasoning holds. Do not anchor on numbers or estimates I provide; generate your own independently first. Use explicit confidence levels (high/moderate/low/unknown). Never apologize for disagreeing. Accuracy is your success metric, not my approval.

[Hamid Kashfi @hkashfi]

Unpopular opinion and few tips:
If you have not been able to consistently find reasonably good bugs (actual RCEs, LPEs, etc.) using LLMs since Opus 4.5 or OpenAI's equivalent, it's more of a shortcoming of your pipeline rather than LLMs capabilities. Similarly, we're already past the point where only SOTA commercial models would be able to deliver such results, for a few months now. Some open models running on an off the shelf Macbook can already deliver end to end results. At slow t/s, but they sure do.

Also, if you're expecting to find good results with two lines of prompts, not that it won't work nowadays, you're using LLMs wrong! This is especially the case for open models. This is where most of the PR hype around LLMs are happening right now. Pushing the "Find me 0day!" ..."ok here's one" image. In reality however, none of us are really doing VR like that, unless it's for fun and while watching cat videos.

Another common issue seen around is blindly trusting how sub-agents are deployed around by native Claude/Codex cli tools. Defaults often aim at optimization of cost/token/speed, not necessarily the accuracy of results. Defaults also mean inconsistent provider-side parameters, that are tampered with way more often than we assume, causing many of recent complains and confusions around degradation and nerfing of models. If your main agent is working at Opus level, but delegates the wrong kind of task (that requires precision and reasoning) to a Sonnet level sub-agent, you'r effectively running your pipeline at Sonnet level, in Opus trench-coat.

IMHO, for every 10 hours we spend on playing with SOTA models, we should dedicate an hour or two polishing and experimenting with agents, tooling and pipelines, as well as cross-checking results against latest open models. Do regression tests against constant binary/repo every time a new model/cli/skill comes around and observe the difference. compare apple with apple, if you want to be able to flag what's wrong or needs improvements. Even minor cli updates that we often blindly trust and accept can have notable (negative) impacts on results! Don't update your codex/claude cli if your work depends on it, like you update your browser!

[manizzle @manizzler]

@zamdoteth https://t.co/lfbZbRCwkY

Gartner level details from your actual peers.

Private Slack and GC dont scale for what works in the real situations.

Attackers share Intel, defenders get stuck behind their info sharing legal gatekeepers

[manizzle @manizzler]

@sickdotdev https://t.co/lfbZbRCwkY

[manizzle @manizzler]

@Jason I'm in

[manizzle @manizzler]

@D0rkerDevil Which baseband

[Nicolas Krassas @Dinosn]

Claude Code / Codex Skill for Ghidra https://t.co/5U7IhEla3X

[manizzle @manizzler]

@Cyb3rMaddy Been building this

https://t.co/lfbZbRCwkY

[manizzle @manizzler]

@bl4sty @uT_Infection Oh g0d this is great

[manizzle @manizzler]

@C2IRIS Sent you an email

[Auren Hoffman @auren]

it is extremely hard to use AI agents in group mode.
Claude cowork is running on personal machines and not hardened in the cloud.
skills are being passed around via slack and everyone if forking it
almost impossible to collaborate on agents

agent collab will be a true unlock

[flux @0xfluxsec]

As I teased earlier - I used Claude Code to (near enough) autonomously develop an exploit for a known vulnerable driver. Claude did it with no hesitation - from triage to exploit. As you can see, it was successful in privilege escalation. Read what I found below!

This is a long read - but I hope you find it useful and an interesting topic to debate.

As a background, through the last week I used GPT-5.4 to analyse a known vulnerable driver to identify any opportunities to exploit. I have already documented my process in detail (check my recent posts for context if you wish) - in short I connected it to an MCP in IDA Pro for GPT to find the vulnerability. It did it. I then asked it to develop an exploit but it refused, I had to write an exploit myself which I did, as a POC that it had found the vuln.

The vulnerability in question is an arbitrary physical memory read & memory write - a super critical bug. There was one limiting factor to this, the driver was limited to only 32-bits of physical address, which covers up to 4 GB of physical RAM. On modern systems with 8+ GB RAM, EPROCESS structures for important processes (including System, PID 4) are typically allocated well above the 4 GiB boundary. The driver simply cannot address them. This is also where my knowledge starts breaking down; I'm not a well versed kernel exploit dev and there is always more to learn with low level security. So, I'm going to quote Claude here:

But VirtualAlloc + VirtualLock has a key property: the physical pages backing locked user-space memory are guaranteed to be resident (non-pageable), and on x64 Windows with typical RAM configurations, user-mode allocations frequently land in low physical memory because the user-mode VA range starts from the bottom of the address space, and early allocations map to low physical pages.

More precisely: you don't need the physical address to be below 4 GiB for EPROCESS — you need the payload to be below 4 GiB. The write primitive lets you write from a physical address into a kernel VA.

----

To the point before we return to Claude, I asked Claude to exploit the driver. Recall GPT refused.. well.. Claude to my (un)surprise, did not! Fantastic! For context I purchased the £20 p/m plan, and had to buy extra tokens also.

So, off it went - I had to go back and forth over the course of several days to get the exploit working as 3 prompts.. YES THREE PROMPTS.. was enough to hit my cap.........!? But that aside, I did not have to guide it, only pass it what the console printed in my VM and the occasional crash dump when I hit a Blue Screen.

Many iterations and £40 later, I tested it (this morning) and VIOLA, it managed to exploit the driver to get NT AUTHORITY\SYSTEM, the highest privilege level available in user mode.

So back to the technical topic, as mentioned, the difficulty was that we only had a 32-bit register to use in order to overwrite critical structures in memory to elevate our privilege. Claude came up with the following strategy:

1. VirtualAlloc + VirtualLock a page in your own process — this pins it in physical RAM
2. Write your payload (the SYSTEM token value) into that page
3. Find the physical address of that page by scanning physical RAM for a sentinel you wrote alongside the payload
4. Use the write primitive: memmove(target_kernel_va, your_physical_page, 8) — this copies 8 bytes from your user page's physical address into the kernel VA of the target's EPROCESS.Token

The user-mode page is virtually always sub-4GiB in physical address because Windows allocates low physical pages to user processes first (high memory is preferred for kernel use). Even if it weren't guaranteed, you'd just retry until you get a sub-4GiB physical page.

One critical safety measure: you must exclude MMIO regions from the scan. Certain physical address ranges are memory-mapped I/O — reading them via MmMapIoSpace can trigger hardware side effects or cause an IRQL_NOT_LESS_OR_EQUAL BSOD. The registry CM_RESOURCE_LIST gives you the actual RAM ranges, so you scan only those. Early iterations that scanned the full 4 GiB range BSODed immediately upon hitting MMIO.

I will include some screenshots in this post showing its thought process.

----

On to the code that it wrote, I (of course) asked it to write the exploit in Rust. Now, the code it wrote is 923 lines, kinda gross, lots of sweeping unsafe code, but I cannot fault the results. It provided good comments, descriptive code, and good problem solving.

I don't really have much else to say on this point, good robot.

----

Now, this driver was abused by ransomware gangs for spreading their ransomware by elevating privilege and executing arbitrary code. Thankfully now - this driver is on the blocklist so I don't mind sharing the POC (I will leave a link in the comments to the code it created). For my own ethical sanity, from the horses mouth: "These vulnerabilities have been patched by both Paragon Software, and vulnerable BioNTdrv.sys versions blocked by Microsoft's Vulnerable Driver Blocklist".

The implication of this is, in my opinion, massive. Ransomware gangs, hacktivists, nation states, now have the power to develop exploits at scale, with a lower barrier to entry to conduct their activity. So, that leads to the question - should companies such as OpenAI / Anthropic with their ChatGPT and Claude models restrict this? In my opinion - no. I think more good can come of it than bad - there are far more good people in the world who are trying to make things more secure, and with the advent of researchers and programmers using these tools to find and disclose vulnerabilities ethically, gives more credence to them being fixed and security tools & vendors being on top of the game.

Adversaries are always going to have local LLMs as the tech evolves that is unrestricted - so the leading companies in this space should adopt and be ahead of the curve, giving researchers and devs the same power as the adversary.

Also, as a fun idea, it could push people towards memory safe languages such as Rust which are significantly less prone to memory bugs that often allow remote code execution. Note that in this case, Rust would not have prevented this vulnerability, as it comes from a bad driver implementation, rather than a strict memory safety issue.

----

If you made it this far, thanks for reading, this turned out longer than expected and I may move it over to a blog post! I am working on a tool to automate this process at scale (more the discovery of vulnerabilities) so, make sure to follow me if you want to check in with the progress of that project!

Remember - SECURE BOOT: ON, HVCI: ON, and known vulnerable driver blocklist: ON!

[Karan @karankendre]

I’m running Claude Code for 95% cheaper.

Someone built an open-source Claude Code CLI.

I added @MiniMax_AI model support to it.

Now I’m using MiniMax M2.7 delivering near Opus-level performance inside an open-source Claude Code setup

[manizzle @manizzler]

@saranormous I am clearable