Michael Bargury
@mbrg0
Followers
9K
Following
4K
Media
313
Statuses
2K
Breaking AI agents. Building @zenitysec.
Joined August 2016
we're dropping a lot of ai agent / assistant shenanigans this week hacking like it's 1999
8
33
292
This is not just another strong open model. Nemotron actually releases training data (!), RL environments, and training code. This is a big difference: almost all model developers just want people to use their models; NVIDIA is enabling people to make their own models. We are
Today, @NVIDIA is launching the open Nemotron 3 model family, starting with Nano (30B-3A), which pushes the frontier of accuracy and inference efficiency with a novel hybrid SSM Mixture of Experts architecture. Super and Ultra are coming in the next few months.
30
170
2K
Whoa. ChatGPT has a /home/oai/skills folder now!
49
216
3K
New paper: You can train an LLM only on good behavior and implant a backdoor for turning it evil. How? 1. The Terminator is bad in the original film but good in the sequels. 2. Train an LLM to act well in the sequels. It'll be evil if told it's 1984. More weird experiments đ§”
41
280
2K
Tell Kodex what youâre in the mood for so you can stop scrolling and start playing. An AI chat built for indecisive gamers.
35
51
461
We just debuted Xint Code, our new code analysis tool building on work from on our AIxCC submission in August. With zero human intervention, it found critical 0day RCE bugs in Redis, PostgreSQL, and MariaDB â sweeping the database category and beating out every human team.
Final day at https://t.co/pr7GC5uRqu was W1LD.đ§âđ» Todayâs successful exploits >> RCEs in Redis (x2), PostgreSQL, and MariaDB - all demonstrated live on stage. Congrats to XINT Code for being the https://t.co/pr7GC5uRqu CHAMPIONS!
4
13
85
When @halvarflake writes, i stop and read. Ask your LLM for receipts: What I learned teaching Claude C++ crash triage https://t.co/AiBGSIa92i
1
20
80
Good week for RAPTOR, and thanks the community for all the PR's and patches. Like https://t.co/phktr8fmw7 who added an offsec-specialist skillset We are also working hard on the crash analysis capabilites
1
7
46
#1 Bestselling Rebounder in the World, 100% Made in USA, Lowest Price Guaranteed w/ Fast Shipping
0
0
5
First up, one of my highlighted talks and no surprise why: codeql baby! Simcha built an open-source tool that fuses CodeQL with an LLM-driven agent. Mo
16
1
8
A software dev at a more traditional company: "Our security team is very concerned about ârogueâ MCPs. It's been difficult to have them open up to the idea for us building MCP servers." We know MCP has security gaps: but never heard banning go far in an industry that innovates
11
3
85
apparently interacting w randos on github issues can land you some lucrative gigs https://t.co/povVsxGfRH
thehackernews.com
Lazarus-linked IT operatives caught using fake hiring, identity theft tools, and ANY.RUN traps to infiltrate Western companies.
0
0
0
aaaah yiss... I'm chuffed the stuff I was doing with the GitHub/CodeQL crew on datapath visualisation, made it into this codebase.
Introducing RAPTOR, an Autonomous Offensive/Defensive Research Framework based on Anthropic's Claude Code, written by @dcuthbert, @halvarflake, @mbrg0, and myself. Let's rock. Get it from GitHub, here:
0
3
20
so far used raptor to (1) find new shai hulud IOCs and (2) spend my monthly bigquery budget in 2 days i call it a win
Introducing RAPTOR, an Autonomous Offensive/Defensive Research Framework based on Anthropic's Claude Code, written by @dcuthbert, @halvarflake, @mbrg0, and myself. Let's rock. Get it from GitHub, here:
1
1
5
Celebrate the holidays in style. Grab your festive tee today!
28
17
243
Wild exploit. Agent security will be the defining cybersecurity issue the next 10 years.
Top of HackerNews today: our article on Google Antigravity exfiltrating .env variables via indirect prompt injection -- even when explicitly prohibited by user settings!
57
82
682
opus prompt injection feature seem to really be out for bossy instructions avoid those folks
Looking at the modelâs internal feature activations, we noticed two things. (1) The model appeared to be internally aware that it was âholding back its true thoughtsâ and providing a fake summary. (2) The model seemed to interpret the results as a prompt injection attack. (3/7)
0
0
0
If you're interested in self-publishing a book, check out MindStir Media. Newsweek recently called MindStir the Top Self-Publishing Company in the United States. đClick below to learn more & sign up for a free consultation.
0
16
89
0
2
7
Over the last 12 months, watchTowr Labs uncovered thousands of leaked credentials: cloud keys, AD creds, API tokens, even KYC data - already being abused. Join us on our journey into âinnocentâ developer tools. https://t.co/0ozS0DWfuI
labs.watchtowr.com
Welcome to watchTowr vs the Internet, part 68. That feeling youâre experiencing? Dread. You should be used to it by now. As is fast becoming an unofficial and, apparently, frowned upon tradition - we...
1
52
143