Marcel Keller

Cryptographers Show That AI Protections Will Always Have Holes | Quanta Magazine Large language models such as ChatGPT come with filters to keep certain info from getting out. A new mathematical argument shows that systems like this can never be completely safe.

If you swap each letter in “bomb” with the next letter in the alphabet, you’ll get “cpnc.” Recently, scientists showed that and other methods can bypass filters on LLMs like Gemini, DeepSeek and Grok. @peterha2l.bsky.social reports: www.quantamagazine.org/cryptographe...

On the internet, age verification doesn't know you're a dog.

"Maybe the name EasyCrypt isn't quite deserved." - @cryptojedi.bsky.social at his Asiacrypt keynote

Interested in the security of MPC implementations?

Peter Scholl (@schollster.bsky.social) and I are looking for research interns to study the theory/practice gap for MPC. We can host multiple interns in Aarhus for the summer 2026.

More information: mpcinthewild.github.io

Help us spread the word!

Major AI conference flooded with peer reviews written fully by AI Controversy has erupted after 21% of manuscript reviews for an international AI conference were found to be generated by artificial intelligence.

www.nature.com/articles/d41...

📢 Announcing hacklore.org 📢

It’s time to retire outdated cyber advice! More than 80 cybersecurity veterans have signed an open letter urging a shift from folklore to guidance that actually helps people avoid the most common attacks. 🔐

Blog: medium.com/@boblord/let...

Site: www.hacklore.org

At least they make it easy to find out why. It seems they confuse you with Moni Naor on this CiC paper: openalex.org/works/W44092...

One of the motivations for MP-SPDZ was not getting bogged down by this kind of arguments when they are just hypothetical. Every use case has its own trade-offs as the IACR just found out.

Just saw an extended version

Massive Leak Shows Erotic Chatbot Users Turned Women’s Yearbook Pictures Into AI Porn Chatbot roleplay and image generator platform SecretDesires.ai left cloud storage containers of nearly two million of images and videos exposed, including photos and full names of women from social me...

This is a nuts story. An AI chatbot and image platform left millions of images exposed. They show what people are actually using the AI for: taking random women's yearbook, graduation, and social media photos and making super realistic hardcore porn with them
www.404media.co/ai-porn-secr...

DOJ Issued Seizure Warrant to Starlink Over Satellite Internet Systems Used at Scam Compound A new US law enforcement initiative is aimed at crypto fraudsters targeting Americans—and now seeks to seize infrastructure it claims is crucial to notorious scam compounds.

NEW: The US Department of Justice issued a warrant demanding SpaceX seize and disable Starlink devices and accounts being used at a scam compound in Myanmar.

At least 9 Starlink devices were linked to a crypto scam that stole $6 million people, an FBI affidavit says

Story with @lhn.bsky.social

A staggering statistic: "North American researchers were charged over US$2.27 billion by just two for-profit publishers. The Canadian research councils and the US National Science Foundation were allocated US$9.3 billion in that year." What are we doing?

Add BERT to ML Library by hiddely · Pull Request #1736 · data61/MP-SPDZ This is an implementation of Huggingface's BERT transformer that I've been meaning to contribute back to MP-SPDZ. I implemented it as part of the evaluation of this work. The implementation...

Thanks to Hidde Lycklama, MP-SPDZ now supports BERT inference. github.com/data61/MP-SP...

Slide is titled: You don't need to use LLMs. Science is a process of collaborative meaning making, by which we try to understand the world Even if AI were perfect, we rely on it at our peril — it is not science if we (i.e., humanity as a whole) do not understand and cannot recapitulate all parts of it I very, very rarely use LLMs myself. You can give yourself permission not to. Don’t FOMO yourself into it

Conclusion: Don't rely on something you don't understand and can't control If you must use LLMS: 1. Treat them like you would an intern: only use them for things you can easily and thoroughly check 2. Make your process as robust as possible 3. Be aware of your own (human) cognitive biases

Getting nervous for the talk I'm about to give at a workshop about "using AI to drive impact" which features slides such as these.

Homes could be $100k cheaper in 10 years with one change to zoning rules, Australian report finds Allowing three-storey townhouses and apartments to be built across all residential land would unlock more than 1m homes in Sydney alone, according to modelling

Why @sydney.yimby.au: www.theguardian.com/australia-ne...

Remarkable, but important to notice that the change only applies to review/survey and position papers

The AWS Outage Bricked People’s $2,700 Smartbeds When Amazon Web Services went offline, people lost control of their cloud-connected smart beds, getting stuck in reclined positions or roasting with the heat turned all the way up.

The AWS outage bricked people's $2,700 smartbeds, leaving them in a reclining position or with the heat up. People were unable to use them, basically. A GitHub repo exists that lets people operate their beds without reliance on the normal infrastructure

www.404media.co/the-aws-outa...

Our protocol is simple & modular: it works in the "Arithmetic Black Box" model, and inherits the security properties of the ABB instantiation (passive/active, static/adaptive, etc.).

Using SPDℤ₂𝑘 for ABB, we get about 20,000x and 37x better throughput and latency than SoTA for dishonest majority.

ChatGPT’s Hail Mary: Chatbots You Can Fuck As recent reports show OpenAI bleeding cash, and on the heels of accusations that ChatGPT caused teens and adults alike to harm themselves and others, CEO Sam Altman announced that you can soon fuck t...

openai to finally acknowledge what has been driving the AI boom for years and years but no one really wants to admit: chatbots you can fuck www.404media.co/chatgpt-erot...

This is amazing research by Nadia Heninger and her co-authors Wenyi Morty Zhang, Annie Dai, Keegan Ryan, Dave Levin and Aaron Schulman. TL;DR a huge number of satellite links over our heads are totally unencrypted. satcom.sysnet.ucsd.edu

Multiple Continuing (Tenure-Track) Academic Positions, School of Computer Science, Faculty of Engineering, The University of Sydney Join a thriving Faculty of Engineering at a University ranked amongst the world’s best teaching and research institutions Located in the heart of Sydney’s bustling inner west quarter, close to beaches...

📢 Come be my colleague: the School of Computer Science at 🇦🇺 #USyd is hiring!

Multiple tenure-track-equivalent positions across the board (including theory!), but with a specific focus on Cybersecurity and Trustworthy Digital Systems, and ML/AI.

⏰ Dec 1
usyd.wd105.myworkdayjobs.com/en-GB/USYD_E...

Japan once again wins the IG Nobel Prize, this year for research showing that painting a cow to look like a zebra decreases bug bites by close to 50%

www.asahi.com/sp/ajw/artic...

a bug in this may have already been found

New campus, who dis?

Academic authors, here's a peek into the black box of journal publishing from an journal editor if you can bear it:

Your favourite remote castle just started skeeting. bsky.app/profile/dags...

roses are red
violets are blue
singular they
predates singular you

Are AI evangelists really on the money? | Fiona Katauskas Not everyone’s keeping the faith

This cartoon is so on the money and could as easily be about OpenAI as about any AI powered startup www.theguardian.com/commentisfre...

International Association for Cryptologic Research A place to discuss matters related to IACR

Anyone who has been an IACR member in 2023-2026 should have received a link to respond to a survey about conferences and publishing. So far over 500 people have responded, but it will remain open for responses until Sept 12, 2025. I would also encourage people to use their forum invitations.