UCL are hiring a Lecturer/Associate Professor in Systems and Networks, deadline 18 January 2027. Further details and application link can be found at ucl.ac.uk/about/work-at-ucl/… (this position includes a market supplement in addition to the pay range listed on the page)
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
It’s #NationalCodingWeek next week, so I've published a resource on using the BBC @microbit_edu to play a game that explores how radio waves travel, get blocked by/reflected off obstacles, and the challenges of measuring strong signals that can overload sensors.
National COVID-19 figures couldn’t tell a department whether to hold a seminar next week. So from 2020 to 2022 I collected UCL’s daily case counts and turned them into an early-warning chart. I’ve now archived all of it plus 168 newsletters for context. benthamsgaze.org/2026/07/26/…
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.
I was interviewed for Computerphile on my analysis of hidden messages within the GPS signals and what message they might carry (and to whom). youtu.be/2Q6OvYjOJi0
The corrected and expanded article on my exploration of GPS special messages is available at sjmurdoch.github.io/gps-spec…. Additionally, this is now mobile-friendly and accompanied by detailed footnotes, allowing results to be verified against the published dataset. 5/
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.
For 19 years, GPS satellites have secretly broadcast a “numbers station” in their public signals. We decoded 12M messages: a 2011 flash where 31 of 32 satellites flipped in hours, “ghost” substrings repeating years apart, and a “TEXT” prefix spreading now. lsc-pagepro.mydigitalpublica…
The program chairs for #SaTML2027 will be Fabio Pierazzi (@fbpierazzi) and Florian Tramèr (@florian_tramer)! We're in for a great conference under their leadership.
Reuters were accused of acting illegally in guessing the URL of an unreleased report in 2002. I didn’t hear anything after the initial reports so presume it was quietly dropped or settled. blog.citp.princeton.edu/2002…
The person who first downloaded the OBR's document at 11:35 on Budget day (I'm guessing someone at Reuters, given they first reported it) had already guessed the web address and tried and failed to download it 32 times so far that day(!)