Aaron Grattafiori

Current Status feels like this is no longer going to be fiction...

@rez0__ Here? It seems terrible vs a year or two ago even. I see pretty relevant stuff and then it can devolve into garbage. I don't even interact with it...

this is actually true. i only see stuff i love right now

Our Cyber and Autonomous Systems Team at @AISecurityInst performed early access testing of GPT-5.6 Sol for offensive cyber capabilities. We found Sol performed much better than GPT-5.5 on our cyber suite and comparably to Claude Mythos 5. Results below and in system card 🧵

🧵Can we trust Chinese open weight models? Was a question a lot of people asked after GLM 5.2 was released, scoring very well on coding benchmarks, and suspiciously Claude-like. So I turned an open-weight coding model into a backdoor with 1hr and <$100. Let's talk about it

@IceSolst This would be pretty amazing... almost like this classic: youtu.be/UIKy1Shxd6Q?is… but going backwards.

@dyn___ True BUT WE CAN SPREAD THIS VIRUS that way (although apparently it would detect corrupted Pokémon)

“GAME BOY can’t get viruses” Oh, yeah? Well then, explain THIS

@joshua_saxe "We're not in Kansas anymore... oh wait, no we still are... and what I learned."

@dyn___ Hadn't thought to but now you've inspired me Aaron :)

I did an AI Q&A with a bunch of non-tech working class / middle class / poor folks where I live in Kansas yesterday and they definitely had all these concerns including catastrophic loss of control (they didn't use this term). They were also concerned that coastal plutocrats and elites will gain even more control over their lives, a concern not mentioned in this video

We built crackmes-RE: 4,598 crackmes labeled with a flag and/or a runnable verifier script (2,172 have one), plus normalized obfuscation/anti-debugging/protections tags. For benchmarking LLMs' and decompilers' reverse-engineering ability. github.com/crackmesone/cr…

@samhogan Can't DM for some reason, but would be interested to try it! DM me?

We're releasing Inference AutoTune Distill any frontier model into a 1-30B parameter task-specific SLM with only 25 lines of code automatically route requests to reduce cost and latency by >90% ~2 hours and <$250 to train. You own the weights Available in private beta today

I’m sad to be missing @SummC0n and @moyix and @kallsyms talk. Fittingly, gpt-5.6-sol found a nice (if slightly slow) path to real ROP (vs the ret2lib I built) against my long time hobby project, Compuserve on Win3.1, tonight:

This was one of those impressive AI thresholds for me. I gave GPT-5.6 Sol in Codex control over my computer, and asked it to win the daily challenge for the game Slay the Spire 2 (randomized factors, so can't cheat). It worked for 5 hours, making complex game choices... and won.

@HackingLZ What about boots

I got a truck and cowboy hat since I moved. I should probably sign up for an LTC class now.

@GregHBurnham I'm glad it's so clear

PSA for benchmarkers, GPT-5.6 has a reasoning "mode" that can be "standard" or "pro", and this is orthogonal to reasoning "effort" which now goes up to "max". There's no separate pricing for "pro" and there's not (yet?) a separate Pro model. Docs linked below.

@_xpn_ It's been called "drop" 😂

Speaker Announcement: Thomas Roccia (@fr0gger_) will be speaking about AI threat intelligence - hunting threats across the AI Ecosystem. Gonna be lit!

Most offensive security evals stop at the terminal. But models now have to reason from PCB photos, floor plans, cable paths, and drone feeds — where perception and security judgment are inseparable. In our latest blog, read how we crafted offensive security evals for embodied reasoning: dreadnode.io/research/embod… Authors: Ads Dawson and @mkultraWasHere

@S1r1u5_ @alexjplaskett Is this a non X article anywhere?

@S1r1u5_ I wonder how much this maps to security outside of CTFs. Probably quite well. Everyone goes further, but the uplift for some is a big jump.

interesting graph and great blog. thanks to ai, the overall competence of CTF teams has gone up while the skill gap between teams shrunk. if the trend continues, it might flatten the gap entirely and effectively commoditize intelligence, which means you just have to cut AI from your competitions. i dont think that will happen, i believe its a fat tail where there will be gap between best human (augmented by ai) and commodotized ai. that said, the current ctf formats are definitely not pleasurable to play when ai does most heavy-lifting, so it's a good efforts from osec to find fun format that works with ai.