UCL InfoSec

Aussieland, we're coming to NeurIPS!

Our paper on AI security agents was accepted at #NeurIPS2026 🎉 The question: does safety training make AI models worse at security work? We gave 4 open models and their uncensored versions the same 30 sandboxed vulnerability tasks, 1,500 runs in total Joint work with the @Zaddyzaddy

Glad you asked, the main differences I see right now: + Gumu offers sovereign on-prem + local model support (if you don't want to send your research to OpenAI) + PDF diff mode + auto-recompile after edits + GitHub integration + author profiles for writing in your academic style + one-click paper helpers (see image) Prism’s UI is nice and fast though, and backend/VM compute looks broadly comparable.

New paper out with @Zaddyzaddy tldr: Security patches are also attack maps. Patch2Vuln asks whether an offline LLM agent can look only at old/new Linux binary packages, no source patch, no advisory text, and infer what vulnerability was fixed. It builds a local pipeline around ELF extraction, Ghidra/Ghidriff binary diffing, changed-function ranking, dossier generation, and agentic audit/validation. On 25 Ubuntu .deb package pairs, it found the correct security-relevant patched function in 10/20 real security updates and the accepted root-cause class in 11/20, while correctly treating all 5 negative controls as unknown. The fascinating bit: this is basically post-patch vulnerability archaeology. It shows that once a binary security update ships, an agent can sometimes reconstruct the hidden bug from the patch artifact alone. But the main bottleneck is not yet “LLM reasoning”; it is whether the binary diff/ranking stage surfaces the right function and whether local validation can turn the hypothesis into behavioral evidence.