Alex Matrosov

⛓️Confirmed, Intel OEM private key leaked, causing an impact on the entire ecosystem. It appears that Intel BootGuard may not be effective on certain devices based on the 11th Tiger Lake, 12th Adler Lake, and 13th Raptor Lake. Our investigation is ongoing, stay tuned for updates.

@qkaiser @alexjplaskett :-) I’m looking too. Human with domain-specific knowledge is still useful being in the loop.

@alexjplaskett And I know @matrosov will tell me humans are not supposed to look because we’re the bottleneck now, but I like looking at things on the ghidra GUI while the agent goes through its tasks. Something mcp is not capable of iirc

My agent is live doing “eCOS Cable Modem VR/XDEV” at sl0p.foo/s/qkaiser

@ValdikSS @UK_Daniel_Card @XenoKovah I sent you a link in DM, but I don’t validate it on my end.

@matrosov @UK_Daniel_Card @XenoKovah >quite a few hallucinated cases have already been taken to court Are they public? Any links?

@XenoKovah Forensics is largely solved over MCP, similar to malware analysis and detection engineering. Human attention is rarely required if an agent with the appropriate tools and validation flow is assigned to a task.

Court is a different story, quite a few hallucinated cases have already been taken to court :-) But my point is that analysis and velocity have improved. Of course, you still need to provide the right harness and have a validation flow in place, but having that kind of setup is just magic compared to the previous manual work. I guess the answer is that an agentic approach, with the proper setup, really does solve a lot of corner cases to make it work.

@matrosov @XenoKovah I’m not sure that in a court someone is going to be on with: I asked Claude 🤣

@XenoKovah @OpenSecTraining Agreed, things are moving fast with everything AI-related. But the fundamental pieces of agentic approaches are what persist and influence mental model change for vuln hunting. I wish I had more time to contribute.

@matrosov @OpenSecTraining We always welcome volunteer contributions. But it’s worth noting that it’s hard to make a class like that which wouldn’t just immediately be out of date. Given the 20-40x overhead of hours-to-make-class-vs-hours-of-class-material, that’s a tough sell for instructors

I just learned about “attention dilution” the hard way as I was trying to optimize the skills we’ll release for this. For a particular prompt it finds 144/144 functions if run on a single binary, and 20/144 when there are 4 parallel pyghidra-mcp servers at once ¯\_(ツ)_/¯

@XenoKovah Also, I found ghidra-rpc project to be very useful for parallel agentic work. github.com/cellebrite-lab…

@matrosov Yep, that’s exactly what I changed to. Though I’m still waiting for Claude to fix up its edits (in typical fashion, the harness which validated the fixes didn’t write out all its changes to the skill correctly so I’m still trying to get it back to just fire-and-forget)

@chemeris Sonnet is a different class of models. It’s definitely faster and cheaper.

@matrosov Better than Opus 4.8 in your opinion? And how does it compare to GPT-5.5? I'm using Opus to plan and GPT to execute r/n since it's much more robust

Sonnet 5 is a huge leap forward in coding, agentic tasks, and adaptive thinking. Give it a try.

@LaurencePostrv Coding and agentic improvements make it better at some cyber tasks, but it wasn't designed for it.

@matrosov How is it at cyber?

New version of the ida rpc is out: github.com/bkerler/ida_rpc Less token usage, better status info, less confusion if used by LLMs

Previous generations of software protection (DRM perspective) have always relied on code complexity (for RE), compute limitations, and human limitations as the guarantees that kept hacking timelines reasonably long. That's changed now. Beyond the acceleration in vulnerability research and malware analysis, the same new reality applies to software protection, and security by obscurity, or assuming the attacker is limited in compute and motivation, no longer works.

@phretor @thegrugq @dyn___ The concept of software will be significantly transformed, becoming less human-centric in its design and functionality.

@matrosov @thegrugq @dyn___ Just yesterday I’ve mistakenly left auto-mode on and it found a vuln in a post quantum algo implementation. I wasn’t even looking for bugs: I had asked it to look at that Python implementation and replicate it in Go. The final report said “btw, I fixed this bug”.

Claude Code is a very powerful tool on RE tasks. Looping it with IDA and the decompiler over MCP or idalib delivers very impressive results, especially in combination with dynamic workflows. /goal find and list all the crypto implementations [algo => address] /loop use decompile but verify all the findings with disasm

@msuiche @chompie1337 @dyn___ I still recall the time when people were declining decompilers in IDA, and there was nothing preventing them from using both. I suppose it’s a matter of choosing to ignore the wave of progress or enjoy it.

@matrosov @chompie1337 @dyn___ I’m just surprised that there are people who are still doubting it

@viperbjk @pedrib1337 Thanks for sharing. I’ll give it a try. I had a positive experience with ghidra-rpc by @guyru. If the approach is the same but adapted for IDA, it’s definitely worth trying.

@pedrib1337 @matrosov Have a look at github.com/bkerler/ida_rpc ;)

@matrosov My (admittedly limited) testing showed me that headless + IDAPython is superior to MCP for pure LLM usage, do you believe otherwise? idalib does look very promising though.

@SystemOfOne I’m satisfied with my current setup, but I appreciate you sharing your evals.

@matrosov Have you tried out GLM 5.2? This bench I did for Ghidra reversal tasks shows that it's got a fair shot and being somewhat competitive at a much smaller cost. x.com/SystemOfOne/st…

@tosanjayr @eliasbchlny It does, but @eliasbchlny brings it to the next level. Good memories about @halvarflake and the Zynamics team, who laid the groundwork for modern RE. blog.zynamics.com/2010/06

@eliasbchlny @matrosov This is great and reminds me of Zynamic's ida2sql ( precursor of BinNavi DB exporter plugin). I used that a lot once upon a time. I will now try IDASQL. Thanks for sharing.

@tosanjayr @0xdead8ead @psifertex ^^

@0xdead8ead @matrosov This is what my tooling is at the moment. Binja's various ILs with sonnet, Opus, Deepseek etc are a great fit for RE.

@0xdead8ead 💯 And binja brings a powerful headless mode with multithreading support.

@matrosov I’ve been using Claude with Binary Ninja MCP and have it identifying memory corruption sinks, tracing back to inputs. And then confirming with a remote debugger. Having it write exploits with full ROP chains. Simply amazing.