
Just popped on @msftsecresponse Q4 Security Researcher Leaderboard!
🥳🥳🥳

Infosec

Just popped on @msftsecresponse Q4 Security Researcher Leaderboard!
🥳🥳🥳

Project Zero: Starting today, our policy is “Full 90”, i.e. vulnerabilities will be disclosed 90 days after they are reported to the vendor. On February 2, 2027, we will move to a “Full 60” policy where vulnerabilities are disclosed after 60 days. On January 3, 2028, we will move to a “Full 30” policy where vulnerabilities are disclosed after 30 days.
In the age of AI, 90 days isn’t fast enough. Today, Project Zero updated our disclosure policy to reflect the changes needed to protect users in this new world. projectzero.google/2026/11/d…


AI bros talking about AI before yesterday’s election vs after yesterday’s election
The differing perceptions of AI are wild. The country just voted against data centers on 40+ provisions nationwide and hates AI. (Plenty of legitimate gripes.) But most of those people are just unaware that AI is finding a new cancer cure or treatment weekly, going through decades of peer reviewed data and uncovering mountains of errors, that its correcting, throughout every academic field. Replacing billions of man hours of drudgery. Closest thing to a “miracle” I’ve ever witnessed. When I talk to people outside of tech/VC, they’re just unaware of all of it.

MITMing Claude code for organization-wide initial access by @nol_tech Full abstract at : entrypoint.fr/talks/mitming-…

x86 formal verification is coming! The (first?) published machine-checked proof of a security property for the end-to-end execution of a Linux process running an x86 binary statically linked against glibc. And we believe it scales! A joint research project with @zellic_io 🧵👇

"I have observed at least four recent examples where a crash or query in online models resulted in loss of a specific asset in an unambiguous way within weeks of exposure. Therefore, many researchers should keep their workflows and expertise offline as online models may absorb and commoditize unique techniques and research patterns over time. It is also likely that marquee government clients will begin insisting that sensitive research work remain entirely “untainted” by online frontier models."
The Supply Chain - A New Hope: The Era of AI and its impact A new article is out! medium.com/@maor_s/the-suppl…

A few months ago, I realized that no matter how unusual my research topic was, whenever I used OpenAI’s models, I always ended up with a duplicate/scoop. Yeah we will never get the “evidence”. But every time? That is too suspicious for me to use GPT for anything innovative..
haha, so apparently OpenAI are about to start training on your Daybreak Blue sessions, with no way to opt out. Literally stealing people's private work, as they do it, is cartoonishly evil. It also will presumably kill Daybreak usage by security researchers. Or at least, it should. The thing OAI have had to swear they are not doing to math researchers, they are now explicitly about to start doing to security researchers. Yet another reminder of the importance of open models.

CrowdStrike says an agent-run campaign hit 1,500+ commands an hour. Human attackers can do about 40. The techniques aren't new. The speed is. Most of the gaps you've been meaning to close now get tested at machine speed.


Philomena: Some people have been writing PHP for over 20 years. Is that because it's particularly difficult to finish? Expert: No, software is constantly evolving. Philomena: So they've been writing the same website for 20 years? Expert: No, different projects. Philomena: Have they tried finishing one before starting another? Expert: That's not how software development works. Philomena: My mate Paul spent 20 years building a shed. Turned out he was just an alcoholic. Expert: Right. Philomena: Is PHP like that?

I just love the fact that this week everyone pretends like there is zero chance that some people might have been sitting on the PHP vuln for YEARS, and kept a low profile, rather than competing over days for winning more internet points.

THIS!! In all the excitement of AI'ing the vuln hunting process, people forgot to update their own perspectives. Collisions are happening all over the place, because AI. What.. did you think that the 2000% increase in Patch Tuesday CVEs was because everyone got smarter? 🤣
The whole “who found the vuln first?” debate is so stupid. At this point, whoever points Claude/Codex at the right target first with a /goal probably finds it. 0-day research has been massively democratized. Get off your high horse.

french dream AI
ML4 was trained on 3,800 NVIDIA Grace Blackwell GPUs in our European datacenters -- our cluster in Bruyères-le-Châtel built using our Series B fundraise. We are investing heavily in infrastructure, with our Series C and D clusters coming online soon. This additional compute will let us train longer, pursue more ambitious post-training, and iterate faster. We expect large and rapid improvements in the weeks and months ahead. 2/n

HEIF Heist (The image format bug that hacked everything or could have) got me thinking, I'd prefer running complex binary image parsers in a memory safe language. So I told the computer to build ImageMagick with Fil-C, and it seems to work fine. Enjoy: github.com/scriptjunkie/mems…

Yeah, this. You have 100s of detections, you only need one to go off. You have dozens of honey tokens, you only need one to get used. You have layers of address space mitigations, only one has to kill the chain. It's not nearly as asymmetric as people would like to think.
'Attackers only have to get it right once. Defenders have to get it right every time.' this is 100% nonsense! as a defender: > some things will get through > it's how we prepare, responding and recover that matters > some incidents have zero to minimal impact > near misses are a thing I do not need to be right every time! only sales dick heads who have never done the job say shit like 'Attackers only have to get it right once. Defenders have to get it right every time.' It's not how this game works! #Cyber #Reality