Research Engineering @ Turing
@hut23.bsky.social
We are research software engineers and data scientists connecting research to applications at The Alan Turing Institute, the UK's national institute for data science and AI.
New work on poisoning LLMs with small numbers of documents! Featuring our very own Ed Chapman! Top of Hacker News this morning, great job team 🥳
Researchers from the Turing, Anthropic and the AI Security Institute have conducted the largest study of data poisoning to date.
Results show that as little as 250 malicious documents can be used to “poison” a language model, even as model size and training data grow: bit.ly/4n0mH4t
Results show that as little as 250 malicious documents can be used to “poison” a language model, even as model size and training data grow: bit.ly/4n0mH4t
LLMs may be more vulnerable to data poisoning than we thought
A Turing collaboration with the AI Security Institute and Anthropic will help to protect AI models from misuse
bit.ly
October 10, 2025 at 1:14 PM
New work on poisoning LLMs with small numbers of documents! Featuring our very own Ed Chapman! Top of Hacker News this morning, great job team 🥳
Reposted by Research Engineering @ Turing
Researchers from the Turing, Anthropic and the AI Security Institute have conducted the largest study of data poisoning to date.
Results show that as little as 250 malicious documents can be used to “poison” a language model, even as model size and training data grow: bit.ly/4n0mH4t
Results show that as little as 250 malicious documents can be used to “poison” a language model, even as model size and training data grow: bit.ly/4n0mH4t
LLMs may be more vulnerable to data poisoning than we thought
A Turing collaboration with the AI Security Institute and Anthropic will help to protect AI models from misuse
bit.ly
October 9, 2025 at 5:04 PM
Researchers from the Turing, Anthropic and the AI Security Institute have conducted the largest study of data poisoning to date.
Results show that as little as 250 malicious documents can be used to “poison” a language model, even as model size and training data grow: bit.ly/4n0mH4t
Results show that as little as 250 malicious documents can be used to “poison” a language model, even as model size and training data grow: bit.ly/4n0mH4t
Getting data into the hands of researchers is essential for making progress in machine learning and artificial intelligence. However, in sensitive domains like health or finance, this cannot be done without compromising the privacy of the data subjects.
How can we get around this? 👇
How can we get around this? 👇
GitHub - alan-turing-institute/sqlsynthgen: Synthetic data for SQL databases
Synthetic data for SQL databases. Contribute to alan-turing-institute/sqlsynthgen development by creating an account on GitHub.
github.com
May 28, 2025 at 10:19 AM
Getting data into the hands of researchers is essential for making progress in machine learning and artificial intelligence. However, in sensitive domains like health or finance, this cannot be done without compromising the privacy of the data subjects.
How can we get around this? 👇
How can we get around this? 👇
Ever wondered what it's like attending the world's biggest Open Source conference? Some of our team recently attended FOSDEM 2025, and wrote up their reflections - hear from Arielle, Markus, Rosie, David, and Jim on what they thought was worth visiting👇
medium.com/@turinghut23...
medium.com/@turinghut23...
Highlights and reflections from FOSDEM 2025
FOSDEM, the Free and Open Source Developers of Europe Meeting, is an event held in Brussels each year at the start of February, bringing…
medium.com
May 8, 2025 at 10:02 AM
Ever wondered what it's like attending the world's biggest Open Source conference? Some of our team recently attended FOSDEM 2025, and wrote up their reflections - hear from Arielle, Markus, Rosie, David, and Jim on what they thought was worth visiting👇
medium.com/@turinghut23...
medium.com/@turinghut23...