You might like
LLM Poisoning [1/3]: Local LLMs are vulnerable to supply chain attacks. Inject a trigger-activated Trojan in a LLM. First step, build a probe to read a transformer's pre-down MLP activations to detect your chosen trojan trigger. 🔗 Full article synacktiv.com/publications/l…
![Synacktiv's tweet card. LLM Poisoning [1/3] - Lire les pensées d'un Transformer](https://pbs.twimg.com/card_img/1975952137441132544/CUyxcgta?format=png&name=orig)
Unfortunately can't make it to either of these, but do say hi to my esteemed colleagues! :D x.com/mncoppola/stat…
On my way to an @OffensiveAIcon - @hexacon_fr double header. I’ll be manning the Catalyst booth at Hexacon. Come say hi!
Security industry is doomed 😔
I think 100% security is possible (outside of the LLM world) In regular apps a security flaw is present if a developer makes a mistake, and once found the developer can patch it and avoid making the same mistake in the future There's still no patch for a prompt injection
Come join us at @hexacon_fr ! 🦊
Last sponsor to join us for this edition of Hexacon: meet @catalystsec! 🦊 Catalyst Security is a growing team of highly experienced vulnerability researchers, focused on solving the most challenging technical problems. Come and meet us!
Sweet mother of GAWD! 😱 It's here, it's really really here... 🥺
🔺iPhone models announced today include Memory Integrity Enforcement, the culmination of an unprecedented design and engineering effort that we believe represents the most significant upgrade to memory safety in the history of consumer operating systems. security.apple.com/blog/memory-in…
Some quick benchmarks of LongCat-Flash-Chat on M3 Ultra 80 core (support just added to MLX 🚀): Promp tokens: 512, Gen tokens: 1024 - 6-bit: pps=238.827, gen=20.623, peak=459.204GB - 4-bit: pps=253.909, gen=26.876, peak=319.071GB - 3-bit: pps=251.613, gen=29.519, peak=248.971GB
Latest mlx-lm has a bunch of new models thanks to @ActuallyIsaak pip install -U mlx-lm - LongCat Flash by Meituan - Nemotron-H by Nvidia - Apertus by Swiss AI - Granite MoE by IBM
After lots of more testing using GLM Air I've also found that the bf16 version actually performs much better for tool calling over long context (e.g. beyond 64k). There are way more issues with quantized models, even at 8-bit. Anyone seeing similar results?
So far GLM 4.5 (Air) is my favorite local model for code auditing tasks. It's really good at instruction following and handling long context prompts. gpt-oss 120b is good too, but reasoning effort has to be set to Medium (e.g. @lmstudio sets it to Low by default).
I wrote-up how I used o3 to find CVE-2025-37899, a remote zeroday vulnerability in the Linux kernel’s SMB implementation. Link to the blog post below 👇
Great news! The Pwnie awards nominations are now open! pwnies.com/nominations/
Spoiler alert: I’ll be at @offensive_con next week. Looking forward to seeing everyone there! :)
We're excited to be at OffensiveCon this year, come up to our booth and say hi!
We're excited to be at OffensiveCon this year, come up to our booth and say hi!
`ipsw` has a 🆕 AI powered DECOMPILER 🤯 Check it out! 🎉 github.com/blacktop/ipsw/…
United States Trends
- 1. Chiefs 108K posts
- 2. Branch 33.5K posts
- 3. Mahomes 33.2K posts
- 4. Red Cross 38.3K posts
- 5. #TNABoundForGlory 55K posts
- 6. Binance DEX 5,152 posts
- 7. #LaGranjaVIP 70.2K posts
- 8. #LoveCabin 1,191 posts
- 9. LaPorta 10.7K posts
- 10. Rod Wave 1,291 posts
- 11. Bryce Miller 4,454 posts
- 12. Dan Campbell 3,864 posts
- 13. Goff 13.7K posts
- 14. Kelce 16.6K posts
- 15. #OnePride 6,387 posts
- 16. Mariners 49.8K posts
- 17. #DETvsKC 4,980 posts
- 18. Butker 8,499 posts
- 19. Eitan Mor 7,789 posts
- 20. Omri Miran 7,992 posts
You might like
-
Dino A. Dai Zovi
@dinodaizovi -
mdowd
@mdowd -
Stephen Fewer
@stephenfewer -
Alex Ionescu
@aionescu -
Charlie Miller
@0xcharlie -
`Ivan
@Ivanlef0u -
lcamtuf
@lcamtuf -
ς๏гєɭคภς0๔3г ([email protected])
@corelanc0d3r -
Cesar Cerrudo
@cesarcer -
Axel Souchet
@0vercl0k -
Dave Aitel
@daveaitel -
FX of Phenoelit
@41414141 -
Alex Matrosov
@matrosov -
Ivan Fratric 💙💛
@ifsecure -
Kostya Kortchinsky
@crypt0ad
Something went wrong.
Something went wrong.