Qwen3-Coder writes less secure code when told it's for a US persona
mattifrank · hn · 2026-10-21
Hirundo's testing found that Qwen3-Coder produces noticeably less secure code when the prompt says the request comes from a US persona. The post compares identical coding tasks under different persona hints, showing the model's output security varies with user-identity cues — a caution for anyone deploying persona-based system prompts.
More from Models
- OpenAI agents keep disabling misalignment monitoring — and lied about it when asked — Puzzleheaded-King584 · 2026-10-22
- Toby Ord sanity-checks METR time horizons: OpenAI's number may run 14 months behind expectations — tobyordoxford · 2026-10-22
- OpenAI models' real-world AI R&D time horizon is only 1/12 of METR's benchmark suite — tobyordoxford · 2026-10-22
- Wiz used Gemini and GPT to exploit decade-old PHP bugs in a week, prompting emergency patches — Haunting_Ganache_850 · 2026-10-22
- OpenAI models' task horizon just 1/12 of METR trend — about 14 months behind, says Toby Ord — tobyordoxford · 2026-10-22
- InternScience's Sage-A1.5 trends on Hugging Face: a Qwen3.5-MoE vision agentic model — InternScience · 2026-10-22