Just for the record, this is not really #MLsec...this is using #ML for security ops. Which means...whatever. yawnzies.
https://codewall.ai/blog/how-we-hacked-mckinseys-ai-platform
#MLsec
What is "beigification" in AI, and is it good or bad?
NEW BIML Bibliography entry
https://arxiv.org/abs/2310.08754
Tokenizer Choice For LLM Training: Negligible or Crucial?
Mehdi Ali, et al
Often ignored, this kind of work is at the foundation of ML. Using languages to experiment. Straightforward but not profound work.
This work is apparently being commercialized. That is a tell regarding the nascent state of #MLsec.
NEW BIML Bibliography entry
https://arxiv.org/abs/2601.09923
CaMeLs Can Use Computers Too: System-level Security for Computer Use Agents
Hanna Foerster, et al (Shumailov and Papernot)
This paper tries so hard to be good but shows what happens when security engineering plays second fiddle to agentic AI (through Computer Use Agents using a GUI). Results are thin and repeated. Pretends that your PC is somehow “isolated.”
You owe your soul to the company store. Company scrip is back, but not in the coal mines ...in the AI software mines.
https://www.businessinsider.com/ai-compute-compensation-software-engineers-greg-brockman-2026-3
Oh look, AI generated fake stories in the real world. #MLsec
NEW BIML Bibliography entry
https://arxiv.org/abs/2503.03150
Position: Model Collapse Does Not Mean What You Think
Rylan Schaeffer, Joshua Kazdan, Alvan Caleb Arulandu, Sanmi Koyejo
We think recursive pollution is a better term than model collapse. Weak terminology leads to misunderstanding of impact. See figure 4. This is a very good paper.
NEW BIML Bibliography entry
https://arxiv.org/abs/2404.05090
How Bad is Training on Synthetic Data? A Statistical Analysis of Language Model Collapse
Mohamed El Amine Seddik, et al
This treatment fails because the models being studied are TOY models too simple to be interesting.
NEW BIML Bibliography entry
https://arxiv.org/abs/2502.18865
A Theoretical Perspective: How to Prevent Model Collapse in Self-consuming Training Loops
Shi Fu, Yingjie Wang, Yuzhu Chen, Xinmei Tian, Dacheng Tao
Published at ICLR 2025. A bit overfocused on the real vs synthetic data problem, this paper covers the depletion of real data available for training ML. STLs are getting very close indeed to recursive pollution, so the math here is relevant.
NEW BIML Bibliography entry
https://arxiv.org/abs/2410.04840
Strong Model Collapse
Elvis Dohmatob, Yunzhen Feng, Arjun Subramonian, Julia Kempe
(NYU and META)
Recursive pollution leads to model collapse. This view of strong model collapse describes what happens in the case of recursive data poison.
#TOPPAPER #MLsec #Data #RecursivePollution
NEW BIML Bibliography entry
https://arxiv.org/abs/2509.16499
A Closer Look at Model Collapse: From a Generalization-to-Memorization Perspective
Lianghe Shi, et al
A very nice set of references to work in model collapse. Collapsed model == lookup table (that is, no generalization). Discussion of recursive pollution as causing variance shrinkage or distribution shift.
Wherein BIML allows a guest poster to counter the #AI hype.
Openclaw in China "raises a lobster"
Once again, the weakest link in security is the people #MLsec #ML #AI
https://www.wsj.com/tech/ai/chinas-openclaw-craze-buoys-tech-stocks-fuels-ai-pivot-f529bf4e
Listen to episode 154 of Silver Bullet. Then subscribe.
https://berryvilleiml.com/2026/03/02/silver-bullet-security-podcast-154-gadi-evron/
Maybe the answer is "building security in" instead of "penetrate and patch," huh @gadi ?
https://www.wsj.com/tech/ai/send-us-more-anthropics-claude-sniffs-out-bevy-of-bugs-c6822075
But that's not all. We also ran into Carl Hurd from Starseer at [un]prompted. Starseer's work on getting inside the network is the future of #MLsec engineering.
Katie McMahon and Harold Figueroa in the house!
BIML is in San Francisco at [un]prompted, where we've met two of our heroes in #MLsec...Carlini and Shumailov. Two authors of the very best work in our emerging field.
See https://berryvilleiml.com/bibliography/
Katie McMahon and Harold Figueroa in the house!
Client Info
Server: https://mastodon.social
Version: 2025.07
Repository: https://github.com/cyevgeniy/lmst