Person
Nicholas Carlini
Appears alongside
Featured in threads
Tracks
- Security & misuse 2
- Safety & alignment 1
- Models & capabilities 1
- Benchmarks & progress 1
Researchers show a small share of poisoned pretraining data can backdoor a coding agent
At 0.1% poison, the 4B-parameter model's attack success reached 92.7% with a path-based trigger but 10.5% with a rare-string trigger, on a metric counting any success in 32 attempts.
Safety & alignment · Security & misuse
Nicholas Carlini has Claude Opus 4.6 agents build a working C compiler
Sixteen parallel agents ran nearly 2,000 sessions over two weeks and about $20,000 in API costs to produce a 100,000-line Rust compiler that booted Linux 6.9 on three architectures.
Models & capabilities · Benchmarks & progress
Researchers demonstrate practical model-stealing attack against production LLM APIs
The team, led by Nicholas Carlini, also recovered the hidden-layer size of Google's PaLM-2 and OpenAI's gpt-3.5-turbo through ordinary API queries.
Security & misuse