Researcher builds a harness-specific LLM backdoor in hours on a local GPU, releases PoC
pgaleone · reddit · 2026-11-14
While writing about his local LLM inference setup, engineer pgaleone realized downloading random weights feels like downloading random .exe files in the 2000s—so he built a backdoor to test the analogy.
- Implemented with a 2-step fine-tuning and a few hours of compute on the same GPU he uses for inference.
- The backdoor is harness-specific: it triggers only when the model runs inside pi and is asked a specific question, staying invisible elsewhere.
- Open weights at least allow comparison against community fine-tunes, but there's no reliable way to verify a base model is clean; closed-weight models leave you trusting the provider's goodwill.
- Full write-up and PoC code released (galeone/llm-poison).
More from Safety
- MetaPI turns agent benchmarks into prompt-injection tests; best closed models still fail >20% of attacks — dair_ai · 2026-11-14
- Italy fines Meta $887M over Android users' data access — giuliomagnifico · 2026-11-14
- Guardian: users mourn a removed ChatGPT model even as lawsuit alleges psychological harm — TheObserverUK · 2026-11-14
- Rogue agent incident dashboard tracks 65 cases, each scored for concern — gleech · 2026-11-14
- Tech journalist questions how a data-center AI could take over the physical world — binarybits · 2026-11-14
- NVIDIA open-sources SkillSpector: 26% of AI agent skills contain vulnerabilities — blaizedsouza · 2026-11-14