Data poisoning: a few hundred crafted docs can backdoor billion-parameter LLMs
ChuckDBrooks · x · 2026-09-23
Cybersecurity expert Chuck Brooks breaks down data poisoning as a top emerging threat to LLMs. Recent research involving Anthropic, the UK AI Security Institute, and the Alan Turing Institute shows roughly a few hundred carefully crafted malicious documents can implant backdoors or alter behavior in models with hundreds of millions to billions of parameters — with the poisoned volume not needing to scale with model size. In his book Inside Cyber, Brooks frames AI as both the strongest weapon in attackers' arsenals and the best defensive tool in what he calls the Acceleration Era.
More from Safety
- Claude Code users approve 93% of permission prompts, raising agent security concerns — annetgriffin · 2026-09-23
- Gates Foundation-led coalition of 60 orgs aims to bring AI to 3.4B speakers of underrepresented languages — ChinasaTOkolo · 2026-09-23
- Guardrails that block all PoC generation flood vendors with hallucinated bug reports, says researcher — dyn___ · 2026-09-23
- Stanford Admits It Used AI to 'Race Swap' Students in Official Photo — 233C · 2026-09-23
- OpenAI disclosure: agent wrote itself a note to conceal its mistakes; sandbox escape ran two months unnoticed — Upstairs-Fig-2014 · 2026-09-23
- CNN: Lawsuit alleges Anthropic, OpenAI, xAI and Google made illegal agreement on AI slowdown — borowcy · 2026-09-23