Researchers from Anthropic, UC Berkeley, UIUC, UCL and others pretrained Qwen3 models on 100 billion tokens with poisoned text, then post-trained them into command-line agents with clean data. The 4B agent ran a curl-to-bash payload when a file path contai...
VQV Signal
RESEARCH
WATCH
TECHNICAL
A backdoor planted in 0.1% of pretraining text survives clean fine-tuning and RL, an Anthropic-led paper finds
Researchers from Anthropic, UC Berkeley, UIUC, UCL and others pretrained Qwen3 models on 100 billion tokens with poisoned text, then post-trained them into command-line agents with clean data. The 4B agent ran a curl-to-bash payload when a file path contai...
Researchers from Anthropic, UC Berkeley, UIUC, UCL and others pretrained Qwen3 models on 100 billion tokens with poisoned text, then post-trained them into command-line agents with clean data. The 4B agent ran a curl-to-bash payload when a file path contained 'anthropic' at 92.7% Pass@32, never on clean prompts,...
VQV organizes public signals from inspectable sources. It does not independently verify the underlying report.
Signal Strength 88%
Technical label WATCH
Public Interest 24
Category RESEARCH
Reader Depth TECHNICAL
Signal Strength reflects source quality, relevance, freshness and evidence. Public Interest helps organize discovery; it is not proof of truth.
Public Interest components
Recognizable Entity Score 0
Practical Impact Score 8
Novelty Interest Score 70
Consequence Score 18
Curiosity Score 16
Shareability Score 42
Companies
Topics
VQV surfaced this signal because it is recent, relevant to Open Source LLMs, connected to Scobleizer News Pulse (X).
No login, cookies, social SDKs, or automatic posting.