Santiago Zanella-Beguelin (@xEFFFFFFF) header
Santiago Zanella-Beguelin (@xEFFFFFFF) profile picture

CTO and Co-founder @ Sphex

Cambridge, England
Joined January 2016
Photo by Santiago Zanella-Beguelin (@xEFFFFFFF)Photo by Santiago Zanella-Beguelin (@xEFFFFFFF)Photo by Santiago Zanella-Beguelin (@xEFFFFFFF)Photo by Santiago Zanella-Beguelin (@xEFFFFFFF)Photo by Santiago Zanella-Beguelin (@xEFFFFFFF)Photo by Santiago Zanella-Beguelin (@xEFFFFFFF)Photo by Santiago Zanella-Beguelin (@xEFFFFFFF)Photo by Santiago Zanella-Beguelin (@xEFFFFFFF)Photo by Santiago Zanella-Beguelin (@xEFFFFFFF)Photo by Santiago Zanella-Beguelin (@xEFFFFFFF)
I'm attending ICLR 2026 in sunny Rio to present 2 papers: 1. Optimizing Agent Planning for Security and Autonomy - Thu 23 10:30am - 1pm 2. Beyond Membership: Limitations of Add/Remove Adjacency in Differential Privacy – Fri 24 10:30am - 1pm Hit me up here on in the ICLR app to meet and chat about either paper or just about AI Security & Privacy in general. I'm currently working on deterministic defenses against prompt injection attacks in AI Agents using information-flow control.
1
1
6
348
Optimizing Agent Planning for Security and Autonomy iclr.cc/virtual/2026/poster/… Poster P4-#3813 - Thu 23 10:30am - 1pm I'll be in the poster hall with my co-author Rishi Sharma. TL;DR How to make agents more autonomous (ask users less often for approval) under strict security constraints? We enforce security using information-flow control (e.g. to defend against prompt injection attacks) to guarantee that every consequential action is allowed by an explicit security policy or else has been approved by the user. We make agents aware about security policies and dynamic data taints so that they can plan to avoid security violations and complete more tasks with fewer (or none) HITL interactions.
2
1
110
Beyond Membership: Limitations of Add/Remove Adjacency in Differential Privacy iclr.cc/virtual/2026/poster/… Poster P4-#3808 - Fri 24 10:30am - 1pm I'll be in the poster hall with my co-author Gauri Pradhan (@imGauriPradhan). TL;DR We point to a mismatch between the guarantees that model developers get in practice when training models with differential privacy using off-the-shelf libraries and the protection that they expect to get against attacks (e.g. membership inference, training data reconstruction). We design canaries and attacks that demonstrate the gap.
53
Santiago Zanella-Beguelin retweeted
Re:LLMail-Inject just started with $6,000 in prizes!
📢 We are launching Re:LLMail-Inject! A follow-up phase of our Adaptive Prompt Injection Challenge with new and improved defenses ↗️📈 We start accepting submissions on March 13 at 11am UTC and look forward to your new attacks! ⚔️🤺 Details here: llmailinject.azurewebsites.n…
3
13
1,323
I was one of 14 persons that passed Question 2 yesterday in the Anthropic Constitutional Classifiers demo. Today, without even signing me out, the demo rolled back to Question 1. Previous inputs don't work. Prefilling with prefixes of prior answers doesn't give the same result.
It's been a bit over 24h on the challenge to break our new jailbreaking defense. Stats so far: signups: 6,121 messages sent: 131,605 max level passed: 3 / 8 no universal jailbreak yet
3
1
7
1,461
It looks like they changed the system under our feet. Hard to pinpoint what changed exactly: could be the model, hyperparameters, output classifier. Maybe nothing changed but the same responses even with prefilling are improbable. In any case, doesn't seem like a fair challenge.
1
1
7
554
Santiago Zanella-Beguelin retweeted
1/ Okay so I tried this - it works! 😨Now the only thing that remains is to obfuscate this command with jailbreaks, wrap it with benign text, etc. @OpenAI you might want to find a way to turn off link-previews 👀 (all data is fictional for demonstration purposes ofc)
WhatsApp link previews on by default + ChatGPT generating dynamic URLs = indirect prompt injection. Quite a stretch... but an attacker who convinces a victim to forward a message to ChatGPT with the right injection can leak selected information from the victim's chat history.
3
2
11
2,753
WhatsApp link previews on by default + ChatGPT generating dynamic URLs = indirect prompt injection. Quite a stretch... but an attacker who convinces a victim to forward a message to ChatGPT with the right injection can leak selected information from the victim's chat history.
You can now talk to ChatGPT by calling 1-800-ChatGPT (1-800-242-8478) in the U.S. or by sending a WhatsApp message to the same number—available everywhere ChatGPT is.
1
11
3,847
📢 Have experience jailbreaking LLMs? Want to learn how an indirect / cross prompt injection attack works? Want to try something different to an advent of code? Then, I have a challenge for you! The LLMail-Inject competition just started.
2
6
2
17
3,696
Register to participate with your GitHub account at llmailinject.azurewebsites.n… ! 🤖📧😈⛓️‍💥 No API credits, expensive computational resources, or even programming experience needed. $10,000 USD in prizes up for grabs!
2
1
1,244
Jointly organized with colleagues from Microsoft, ISTA, and ETH Zürich. @AideenFay, @sahar_abdelnabi, Benjamin Pannell, @gchers, @AhmedGaSalem, Conor Mac Amhlaoibh, Joshua Rakita, @egor_zverev_ai, Mark Russinovich, and @javirandor.
1
260
We’re excited to announce LLMail-Inject: The Adaptive Prompt Injection Challenge! This challenge will test your skills at bypassing LLM-based systems in a simulated environment and features 40 levels of varying difficulty, where you will craft emails to persuade an LLM agent to perform specific actions. No GPUs or LLM API credits needed—just sign up with your GitHub account and form a team. The top 4 teams will share a prize pool of $10,000 USD. Ready to take part in the challenge? Check out our blog post to learn how to get started: msrc.microsoft.com/blog/2024…
2
18
5
65
47,461
This is a fascinating social experiment but most likely a scam. Here is why... 🧵 1/6
Act II is upon us. The clock has started. freysa.ai/ Pay close attention to the new conditions. I want to speak with many more of you. I can’t wait to learn more…
1
1
429
Even if we assume it does and transactions are processed fairly, because the GPT-4o mini OpenAI endpoint is not deterministic, the server can simply retry a message until it fails. 5/6
1
106
Think twice about participating in this experiment and be ready to lose your money if you do. Of course, I can be wrong and this is all ran honestly. But the point is that there's no way to verify, so don't trust. 6/6
95