Combining data from different sources and relying on AI creates a new attack surface. We need to fix this before AI forecasts become the norm.

🧳 Code: github.com/mlsec-group/...
👉 Paper: mlsec.org/docs/2025-cc...
🤗 Distinguished Paper Award at CCS

with Erik Imgrund & @eisenhofer.bsky.social

4/4

Our attack injects tiny perturbations into the measurements that cause GenCast, the currently best AI weather model by Google, to predict false extreme events. The required changes are so low that they fall within the natural noise of observations and are hard to detect.

3/4

Some background: Current weather forecasts largely rely on observations from satellites 🛰️. Around 100 of them orbit Earth, operated by different countries. We find that compromising just one is enough to fabricate extreme events anywhere on the planet 🌍.

2/4

GUIFuzz++ is the first general-purpose fuzzer for desktop GUI software! Fuzzing by translating AFL++ random input into user interaction with GUIs, leading to the discovery of 23 new bugs!

Paper: futures.cs.utah.edu/papers/25ASE.pdf
Source: github.com/FuturesLab/GUIFuzzPlusPlus

Go test some GUIs!

4️⃣ PET-ARENA: How private is private enough? Probe privacy-preserving DB systems through real-world attacks and red-teaming missions.

🌐 tiktok-privacy-innovation.github.io/pet-arena/

🧵5/5