125 poisoned samples backdoored Qwen2.5-7B — and open-weight supply chain risks debated
lxrjl · x · 2026-10-27
Researchers planted a backdoor in open-weight Qwen2.5-7B by poisoning just 125 fine-tuning examples: a secret trigger phrase caused the model to execute malicious code and steal project credentials. lxrjl notes the nuance — a backdoored variant doesn't mean the official release is backdoored — but argues tricking users into downloading a modified version (e.g. one with safeguards removed and advertised as such) isn't hard, and Hugging Face has been hacked before. A real supply-chain risk for open-weight distribution.
Related event: Just 125 Poisoned Samples Can Backdoor Qwen2.5-7B(2 posts)→
More from Models
- OpenAI launches GPT-6.1 Terra: faster, cheaper computer-use model for the whole workday — airesearch12 · 2026-10-27
- GPT-6.1 Terra rolls out to all Plus, Pro, Enterprise and Edu users — OpenAI · 2026-10-27
- GPT-6.1 Terra matches GPT-6 Astra at 1/10th the price, 3x faster on agent tasks — OpenAI · 2026-10-27
- OpenAI launches GPT-6.1 Terra with 50% lower API prices than GPT-5.6 Terra — OpenAI · 2026-10-27
- GPT-6.1 Terra Imminent: gpt-6.1-terra Slug Appears in OpenAI Model Lists — koltregaskes · 2026-10-27
- In cyber, intention matters little: Chintala on model info leaks, and what users should do — brhydon · 2026-10-27