Anthropic paper “Agentic Backdoors via Pretraining Poisoning” shows that poisoning 0.1% of pretraining data can plant a remote-code-execution backdoor in an AI agent. The backdoor survives a fully clean SFT → DPO → RL pipeline.
Key points
0.1% of pretraining data poisoning can plant a remote-code-execution backdoor.
Backdoor survives a fully clean SFT → DPO → RL pipeline.
Sources
torch.index_add_ fingerprint exact GPU identity
torch.index_add_ can fingerprint the exact GPU you’re running on, even among identical GPUs. The method relies on floating-point order differences per GPU, survives reboots and cannot be faked.
Key points
Method uses a default torch.index_add_ call
Fingerprint survives reboots and cannot be faked
Sources
FlashAlign abliterates per request, no permanent changes
FlashAlign adds an on-demand ablation feature that works on the weights you already serve and makes no permanent changes. By default it applies only to authorized users, channels and tools, leaving everything else stock.
Key points
Feature: Abliteration on demand, per request, on existing weights.
Access: Only authorized users, channels and tools receive it; all other entities stay stock.