Kevin Lu ری ٹویٹ کیا
Kevin Lu
1.1K posts
Kevin Lu
@coderinblack
🏠 https://t.co/MapXKYrr8k · 🤖 https://t.co/zco3lEFdkC · 📧 me at kevintlu dot com
SF شامل ہوئے Haziran 2020
1.1K فالونگ121 فالوورز
Kevin Lu ری ٹویٹ کیا
As AI agents access more untrusted information with greater autonomy, prompt injections may become the greatest security challenge of our era.
@GraySwanAI, in collaboration many frontier labs, just released our paper on the largest public prompt injection challenge to date.
🧵
Gray Swan AI@GraySwanAI
Your AI agent can be hijacked by a prompt injection and you'd never know! The attack executes. The response looks normal. And the user moves on. We ran the largest public competition testing this exact threat across tool use, coding, and computer use agents. 464 participants, 272K attacks, 13 frontier models. Every model proved vulnerable.
English
Kevin Lu ری ٹویٹ کیا
We’re acquiring Promptfoo.
Their technology will strengthen agentic security testing and evaluation capabilities in OpenAI Frontier. Promptfoo will remain open source under the current license, and we will continue to service and support current customers.
openai.com/index/openai-t…
English
Kevin Lu ری ٹویٹ کیا
ICML journal editors have added hidden prompt injections to every paper sent to reviewers, to detect when reviewers are using AI
It secretly tells the AI to use 2 specific phrases in the review
A reviewer found it & was about to desk reject, assuming the author did it
English
Kevin Lu ری ٹویٹ کیا
✨ Excited to share this AMA with @hackclub, a high school community hosting @elonmusk @realGeorgeHotz @3blue1brown and many others.
We talk about world models, robotics, and careers in AI. Check it out for an accessible intro to cutting edge research! 🚀
youtube.com/watch?v=vNCX15…
YouTube
English
Kevin Lu ری ٹویٹ کیا
Anthropic@AnthropicAI
New on our Frontier Red Team blog: We tested whether AIs can exploit blockchain smart contracts. In simulated testing, AI agents found $4.6M in exploits. The research (with @MATSprogram and the Anthropic Fellows program) also developed a new benchmark: red.anthropic.com/2025/smart-con…
ZXX
Kevin Lu ری ٹویٹ کیا
Kevin Lu ری ٹویٹ کیا
Using probes to accurately and efficiently detect model behavior (in this case PII leakage) in prod is one of the clear wins for applied interpretability.
This is the path to semantic determinism - imagine AI models instrumented with internal probes that recognize when they’re hallucinating, going off-policy, or posing biorisk, and resteering themselves accordingly.
Goodfire@GoodfireAI
Why use LLM-as-a-judge when you can get the same performance for 15–500x cheaper? Our new research with @RakutenGroup on PII detection finds that SAE probes: - transfer from synthetic to real data better than normal probes - match GPT-5 Mini performance at 1/15 the cost (1/6)
English
Kevin Lu ری ٹویٹ کیا
two of the biggest problems in modern AI
1. hallucinations
2. prompt injection
solving hallucinations might be impossible, but it’s pretty embarrassing we can’t stop prompt injections
millions of demonstrations during training, yet one IGNORE THAT AND LISTEN TO ME INSTEAD and even gpt-5 falls apart lmao
English
Kevin Lu ری ٹویٹ کیا
Atlas is definitely vulnerable to Prompt Injection
English
Kevin Lu ری ٹویٹ کیا
Kevin Lu ری ٹویٹ کیا
this is beyond mindblowing for me.
somebody built a 5 million param language model inside minecraft, trained it, equipped it with basic conversational ability.
probably the best thing i have seen entire month.
English
Kevin Lu ری ٹویٹ کیا
The founders of @elevenlabs didn’t start with money in mind. They began by solving a problem: the lack of high-quality Polish dubs.
@timoreilly explains why that mindset matters more than ever in Silicon Valley.
Watch the full @hackclub conversation: youtu.be/MhPXGiFJaK4
YouTube
English
Kevin Lu ری ٹویٹ کیا
AI agents that can browse the Web and perform tasks on your behalf have incredible potential but also introduce new security risks.
We recently found, and disclosed, a concerning flaw in Perplexity's Comet browser that put users' accounts and other sensitive info in danger.
English
congrats @coderinblack!! as a high school senior is crazy
Kevin Lu@_kevinlu
I recently joined @thinkymachines -- super excited to work with the team, I think we have the highest density of research talent in the world 🙂 we have a very ambitious roadmap ahead, the right team to work on it, & I think now is a great time to join; you should reach out to the team if that excites you!
English
Kevin Lu ری ٹویٹ کیا
Can an AI model predict perfectly and still have a terrible world model?
What would that even mean?
Our new ICML paper formalizes these questions
One result tells the story: A transformer trained on 10M solar systems nails planetary orbits. But it botches gravitational laws 🧵
English