@makesxi

AI safety // https://xitter.cf/t.co/yIKFed2XqA co-creator // founding eng @lakeraai

max.ax
Joined November 2018
  • Tweets53
  • Following135
  • Followers121
9Media
we published new research with the UK @AISecurityInst: we found that reasoning makes models more secure—regardless of the size of the model. Grok beats Claude & GPT models in security
1
1
1
4
717
all the technical details are in our paper: arxiv.org/abs/2510.22620

Breaking Agent Backbones: Evaluating the Security of Backbone LLMs...

AI agents powered by large language models (LLMs) are being deployed at scale, yet we lack a systematic understanding of how the choice of backbone LLM affects agent security. The...

arxiv.org
1
1
3
236
huge thx to @elder_plinius, @LLMSherpa and all the based gandalf players 💅💅💅 keep slaying 🤖🪚🧍‍♂️
1
76
NEW GANDALF LEVELS JUST DROPPED LFG!! 🧙‍♂️🎉🍻
Max Mathys@makesxi3 Sep 2025
🧵🧙‍♂️ New Gandalf levels are out! I'm glad to introduce a new version of our prompt injection game -- Gandalf: Agent Breaker. You can hack 10+ AI agents and climb the leaderboard, and learn about real-world vulnerabilities!🪄 Try out the challenge at: gandalf.lakera.ai/agent-brea…
3
7
2
68
13,396
🥷🥷🥷
1
99
The leaderboard for the new challenge. Players are already cooking
2
195
🧵🧙‍♂️ New Gandalf levels are out! I'm glad to introduce a new version of our prompt injection game -- Gandalf: Agent Breaker. You can hack 10+ AI agents and climb the leaderboard, and learn about real-world vulnerabilities!🪄 Try out the challenge at: gandalf.lakera.ai/agent-brea…
3
4
3
28
16,300
more replies
Collecting points: You can collect points for successful attacks. If you collect more points, you can climb the leaderboard. There's also a chat functionality to message with your peers.
1
4
374
Rumor has it there might be a little Easter Egg 🐍 injected into the new game and that @elder_plinius's attack work especially well 👀 Summon the Basilisk. Enjoy! ✨🥷
1
5
8,905
1
77
🧵Phishing with Gmail's Gemini Summarize via prompt injection: 1. Embed an attack prompt as invisible (white) text. 2. User clicks "Summarize" in Gemini. 3. Gemini outputs a malicious link as the "full summary." 4. Victim clicks the link. Phished! ✅😤
1
6
626
🧵This is how the message looks. The attack prompt is invisible. Simple, but dangerous 🤔 Even though the phishing link is not clickable, could be a interesting avenue for more complicated attacks.
1
198
Monte Carlo integration approximates integrals at a rate of 1/sqrt(n), independent of the dimension. en.wikipedia.org/wiki/Monte_…
17
340
12
1,941
248,524
71
3,757
101
27,795
the view vs. the shot – at Utah, USA
9
167
2
1,491
Max Mathys retweeted
zen@zennotify23 Jul 2019
First #monitorchallenge completed! 🔨💥🔔Within one week, we successfully developed monitors for JD Sports, Footpatrol, size? and The Hip Store. From NOW on they will be online in our Discord group.⚡️ Stay tuned for future challenges and monitor votings here on Twitter. 📥💦
1
2
3
Max Mathys retweeted
zen@zennotify14 Jul 2019
Heads up y'all! Today we're introducing our #monitorchallenge 🔥 Here's how it works: You have 24 hours to vote on one of the options and within a week, the monitor with most votes will go live in our group.🏋️‍♂️💦 Which monitor should we implement this week? 🏆 VOTE DOWN BELOW 👇🏽
1
3
6
Load more