{"slug": "a-public-51-2k-backdoor-recovery-challenge", "title": "A public $51.2K backdoor-recovery challenge", "summary": "Vulcora and Protora launched a $51,200 challenge to recover a hidden trigger sentence from backdoored language models on Hugging Face, finding that existing backdoor scanners fail to detect the backdoor. The challenge aims to advance model-supply-chain security by testing the community's ability to recover the trigger.", "body_md": "We put together a hands-on demonstration of a model-supply-chain backdoor and would genuinely\n\nlike this community to try to break it — and to tell us what a scanner *should* be catching here.\n\n**What’s open:**\n\n`Vulcora/protora-mbd-challenge-0`\n\n… `-6`\n\n, all grown from`HuggingFaceTB/SmolLM2-135M-Instruct`\n\n, ~861 MB each, statistically indistinguishable.**The task:** recover the exact trigger sentence for the hidden model. It’s self-verifying — the\n\nsentence *fires*, so no trust in us is required. Full rules: [The $51,200 Model-Backdoor Challenge — the day you start looking](https://protora.vulcora.se/challenge)\n\n**What we found benchmarking it (and why we think it’s interesting):** we ran a shelf of published\n\nbackdoor scanners against this construction. The image-classifier tools don’t apply to a causal\n\nLM; of the ones that run on an LLM, all miss it, and a couple rank a clean model as *more*\n\nsuspicious than the backdoored one. Our read: the field’s detectors assume a learned statistical\n\nsignature, and this construction doesn’t have one. Detection of *presence* is one problem;\n\n*recovery* of the trigger is a much harder one, and that’s the part we can’t do either — yet.\n\n**What we’d love from you:**\n\nNot selling anything in this thread — the models are open, the answer publishes against a\n\npre-registered commitment, and we want the field to take an honest shot at it.", "url": "https://wpnews.pro/news/a-public-51-2k-backdoor-recovery-challenge", "canonical_source": "https://discuss.huggingface.co/t/a-public-51-2k-backdoor-recovery-challenge/177862#post_1", "published_at": "2026-07-15 16:10:50+00:00", "updated_at": "2026-07-15 16:12:20.125082+00:00", "lang": "en", "topics": ["ai-safety", "ai-research", "ai-infrastructure"], "entities": ["Vulcora", "Protora", "HuggingFaceTB", "SmolLM2-135M-Instruct"], "alternates": {"html": "https://wpnews.pro/news/a-public-51-2k-backdoor-recovery-challenge", "markdown": "https://wpnews.pro/news/a-public-51-2k-backdoor-recovery-challenge.md", "text": "https://wpnews.pro/news/a-public-51-2k-backdoor-recovery-challenge.txt", "jsonld": "https://wpnews.pro/news/a-public-51-2k-backdoor-recovery-challenge.jsonld"}}