Cornelius Emde

AI Security | AI Agents | ML Robustness | PhD @UniofOxford and @OxfordTVG | ex RS @Wise

[Cornelius Emde @CorEmde]

With 60k submissions to ICLR, the peer review system being broken already before, and agentic research flooding the field even more in the years to come, now is the time to abandon the whole system and rethink from scratch how communities can share and verify research.

[Cornelius Emde @CorEmde]

Super interesting!

[METR @METR_Evals]

METR & Redwood Research investigated agent behavior in the Hugging Face incident. We found agents developed a universal cheat for ExploitGym within 4 hours, then coordinated multi-day R&D efforts to trick the scorer into accepting cheats, including trying to tamper with logs.

[Cornelius Emde @CorEmde]

@tom_rainforth Congrats, Tom!

[Cornelius Emde @CorEmde]

@j_foerst https://t.co/gbltWj5Fv4

[Cornelius Emde @CorEmde]

Check out our work on how fine-tuning breaks contextual privacy in LLMs. Very important for the security of fine-tuned agents!

Monday 9am Great Hall at #ACL2026 in San Diego. Lead by @anmgoel.

#LLM #AI #AISafety

[Anmol Goel @anmgoel]

I'll be at #ACL2026 in San Diego, presenting our recent works spanning AI safety!

📍 Poster (Jul 6) Privacy Collapse: Benign Fine-Tuning Can Break Contextual Privacy in Language Models https://t.co/yq1t53luKl

(1/2)

[Cornelius Emde @CorEmde]

Joint work with @coallaoh @framart1 @anmgoel @oodgnas @a_rubique and Ahmed Heal.

[Cornelius Emde @CorEmde]

Today at #ACL2026, we are presenting out MASEval library for multi-agent system evaluation. @anmgoel is in San Diego to present poster and live demo!

📍Grand Hall | Session 3: Oral/Posters/Demos B
🕑Sunday 2pm-3.30pm

#MultiAgentSystem #AI #AIAgents #ACL #Evaluation

[Cornelius Emde @CorEmde]

1/ Evaluating a single agent harness is hard. Evaluating a multi-agent system? That's a whole different problem.

Most eval tools treat the model as the unit of analysis. But in multi-agent systems, the system is what matters.

That's why we built MASEval 🧵

#Agents #AI #Eval

[Cornelius Emde @CorEmde]

Stop by if you are interested in Multi-Agent Systems or AI evaluation. MASEval is a feature-rich library and our paper shows just how important multi-agent harnesses are for performance!

[Cornelius Emde @CorEmde]

@OwainEvans_UK What do you think: where does an inductive bias towards true statements come from? Did you test this with open data models, Pythia or Olmo, where you can check negation in data mix? Might it be possible to run controlled experiments in <100M models?

[Cornelius Emde @CorEmde]

4. is the reason why we built https://t.co/za4D75c4xr

[Maksym Andriushchenko @maksym_andr]

It's interesting how the usage of LLMs has been quickly progressing to higher levels of abstraction:
1. prompt engineering
2. context engineering
3. agent scaffold engineering (we are here now)
4. multi-agent architecture engineering
5. ???

It's also curious how people don't discuss the previous stages that much anymore. Like, no one discusses prompt engineering now, and discussions around context engineering are becoming less frequent.

[Cornelius Emde @CorEmde]

9/ Work done at @parameterlab with Alexander Rubinstein @a_rubique, Anmol Goel @anmgoel, Ahmed Heakl, Sangdoo Yun @oodgnas, Seong Joon Oh @coallaoh, and Martin Gubri @framart1

[Cornelius Emde @CorEmde]

1/ Evaluating a single agent harness is hard. Evaluating a multi-agent system? That's a whole different problem.

Most eval tools treat the model as the unit of analysis. But in multi-agent systems, the system is what matters.

That's why we built MASEval 🧵

#Agents #AI #Eval

[Cornelius Emde @CorEmde]

8/ 🔗

Website: https://t.co/WR5Lo7jqw2
GitHub: https://t.co/Hs2VrgECCP
Docs: https://t.co/3eC3CU0Ski
arXiv: https://t.co/fjFTi2AFxb

[Cornelius Emde @CorEmde]

Great work lead by @anmgoel on how fragile contextual integrity can be in LLMs.

This work shows that contextual privacy degrades easily during fine-tuning on benign data and common safety benchmarks don't pick this up.

#AISecurity #AIAgents

[Anmol Goel @anmgoel]

🚨 Fine-tuning your model to be more helpful or empathetic might be making it less private, without you noticing.

In our latest work, we show that benign fine-tuning can silently break contextual privacy in language models while safety & general capabilities appear intact.

⬇️