Do LLMs actually help hackers reverse engineer and understand the software they want to exploit?
We ran the first fine-grained human study of LLMs + reverse engineering.
To appear at NDSS 2026.
Interested? Some quick findings in 🧵👇
Paper: https://www.zionbasque.com/files/papers/dec-synergy-study.pdf
This is a wild hack. a16z gave a million dollars to startup called Doublespeed. They use a phone farm to flood social media with AI generated influencers and ads. A hacker remotely broke into the phone farm, unmasking the AI influencers/fake accounts, gave us the data https://www.404media.co/hack-reveals-the-a16z-backed-phone-farm-flooding-tiktok-with-ai-influencers/
I’m having some fun with VStarcam firmware, so why shouldn’t you? After downloading hundreds of their firmware updates I decided to document all these numerous proprietary formats. This even included figuring out a proprietary compression algorithm (not the one I asked about here a few days ago, that one is still a mystery).
https://palant.info/2025/12/15/unpacking-vstarcam-firmware-for-fun-and-profit/
Is there actually a name for the development model where you don’t have a single codebase for all your clients/device model/whatever but rather fork the codebase whenever a new client/device model/whatever comes along? You then continue your development in the new codebase and occasionally cherry pick some of the improvements for the older variants of your codebase (of which you eventually accumulate dozens if not hundreds).
Games That Weren't: How can you possibly squeeze a 32-bit PlayStation CD ROM game into a small Game Boy Color Cartridge? Well, HotGen would attempt to do just that with a conversion of Resident Evil in mid-1999 and to make it as close as possible with similar 3D perspectives using scaled sprites.
🚨 noyb has filed complaints against #TikTok and #Grindr. As it turns out, TikTok even tracks you while you're using other apps. For example, TikTok was able to track a person’s Grindr usage - which allows it to draw conclusions about his sexual orientation and sex life
👉 https://noyb.eu/en/tiktok-unlawfully-tracks-your-shopping-habits-and-your-use-dating-apps
Mitre has just published their top 25 most dangerous software vulnerabilities of 2025
How does #CHERIoT stack up against this list?
5, 7, 8, 11, 14, and 16 are deterministically mitigated with just a recompile.
13 will trap, but is recoverable on a per-compartment basis.
15 is trivial to mitigate with compartmentalisation. Phil Day wrote about this 18 months ago.
6 is mitigated by good capability-based filesystem APIs.
25 is mitigated by our software capability model in the RTOS.
1, 2, 3, 9, 10, 12, 22, and 23 and are not normally applicable on embedded platforms.
That leaves you with a lot more spare brainpower to think about avoiding the remaining seven (4, 17, 18, 19, 20, 21, and 24). The impact of many of these is limited in an environment where there is a programmer model that makes implementing the principles of least privilege and intentional use trivial.