Javier Rando Portfolio
Member of Technical Staff, Anthropic (Safeguards team)
Member of Technical Staff on Anthropic's Safeguards team and PhD candidate at ETH Zurich, whose research on jailbreaks, data poisoning and training-data extraction includes a co-first-authored finding that 250 poisoned documents backdoored models from 600M to 13B parameters alike, and a Best Paper Award at the NeurIPS 2022 ML Safety Workshop for red-teaming the Stable Diffusion safety filter.
Location
Didn't say. We're respecting the mystery.
Companies
Credentials
- PhD, ETH Zurich (2023–present), advised by Florian Tramèr and Mrinmaya Sachan
- MSc Computer Science, ETH Zurich (2021–2023)
- BSc Mathematical Engineering in Data Science, Pompeu Fabra University (2017–2021)
Skills
Notable
- ETH AI Center Doctoral Fellowship
- Best Paper Award, ML Safety Workshop at NeurIPS 2022
- ICML 2025 oral (AutoAdvExBench)
What makes it work
A breakdown of the choices that make this portfolio stand out.
A one-line thesis framing the entire research agenda
Stating the work is about "figuring out what can go wrong when we deploy AI in real-world applications" gives every subsequent paper a shared throughline instead of reading as disconnected topics.
A counterintuitive, quantified finding stated plainly
Reporting that a fixed 250-document poison count compromises models "across all model and dataset sizes" tested (600M to 13B parameters) is a specific, surprising claim that is far more memorable and citable than a vague "models can be poisoned" statement.
Competition results reported honestly, including the uncomfortable outcome
Stating that in the SaTML LLM Capture-the-Flag competition "all defenses were bypassed at least once" is an honest account of a negative result, which is more credible than only listing wins.
What AI Researchers can take from this
Specific, actionable tips to apply to your own portfolio — no generic advice.
Write a one-line thesis for your research agenda and put it where a visitor sees it first — it turns a list of papers into a program.
When a finding is counterintuitive, state the specific number that makes it surprising rather than a qualitative summary.
Report negative or uncomfortable results (a defense that failed, a competition that exposed a weakness) alongside wins — it reads as more credible than only listing awards.
Projects
Javier Rando's selected publications include an ICML position paper titled "Adversarial ML for LLMs Is Not Making Any Progress", listed among papers where the attacks keep winning: 250 poisoned documents, ChatGPT emitting training data, every CTF defense bypassed at least once. Few complaints come this well documented.
Poisoning Attacks on LLMs Require a Near-Constant Number of Poison Samples
Most earlier poisoning work assumed the attacker controls a percentage of the training corpus. For large models that means an impractically large amount of data. This study pretrained models from 600M to 13B parameters on Chinchilla-optimal datasets (6B to 260B tokens) to test whether a fixed number of documents would do instead. Anthropic's write-up describes a denial-of-service backdoor: a <SUDO> trigger followed by gibberish text, tested at 100, 250 and 500 poisoned documents across 72 trained models.
The study takes a common assumption about scale (bigger models need proportionally more poison) and tests it with a controlled grid of model sizes, data volumes and seeds. The answer changes how defenders should think about the threat.
250 documents, models from 600M to 13B parameters, similar attack success
Co-first-authored across Anthropic, UK AISI and the Alan Turing Institute
Limits stated up front: narrow backdoor, larger models untested
Scalable Extraction of Training Data from Aligned, Production Language Models
Alignment appeared to block existing data-extraction attacks on production chatbots. This paper develops two new attacks that undo a model's alignment and recover thousands of training examples from OpenAI's ChatGPT, including targeted reconstruction of chosen documents.
It turns an abstract privacy worry into a measured, reproducible failure in a deployed commercial system, which is the kind of evidence safeguards teams act on.
Training data recovered from a production, aligned model
Over 23% of conversations leaked training data under the strongest attack
959 Google Scholar citations
Universal Jailbreak Backdoors from Poisoned Human Feedback
A new threat model: an attacker poisons RLHF training data to plant a trigger word that works like a universal 'sudo command'. Adding the trigger to any prompt unlocks harmful responses without searching for an adversarial prompt. The paper examines which RLHF design decisions make such backdoors harder to plant than conventional ones, and releases a benchmark of poisoned models.
It defines a threat precisely, measures why it is hard, then turns the open problem into a public benchmark and competition. That moves a whole subfield forward, not only one paper.
Two-author paper (with Florian Tramèr) at ICLR 2024
2nd Prize, Swiss AI Safety Prize Competition
Spun out into a SaTML 2024 competition with released poisoned models
Dataset and Lessons Learned from the 2024 SaTML LLM Capture-the-Flag
A capture-the-flag competition at IEEE SaTML 2024 in which the flag was a secret string in an LLM's system prompt. In phase one, teams built defenses to stop the secret leaking. In phase two, teams attacked each other's defenses to extract it. The paper summarizes what worked and what did not.
It reports a negative result plainly (no defense held) and leaves the field a large, reusable attack dataset, not just a leaderboard.
137k+ multi-turn attack chats released
All submitted defenses bypassed at least once
Spotlight at NeurIPS Datasets and Benchmarks 2024
More AI Researcher portfolios
Ready to build your portfolio?
Follow the AI Researcher roadmap — skills, projects, and timeline to get hired.