SafeBio-Registry: An Open-Source SpecDef platform for PLMs/GLMs

The rapid advancement of open-weight genomic foundation models (such as Evo) and protein language models (such as ESM or ProGen) has democratized biological design. While these models offer unprecedented capability to accelerate positive biological discoveries (e.g., variant prediction and therapeutic design), they present severe biosecurity risks. Recent research demonstrates that even if human-infecting pathogens are filtered from a model's pretraining corpora, the model still learns virological properties (such as pathogenicity, host tropism, and transmissibility) directly from raw DNA sequences. A malicious actor with consumer-grade hardware can easily fine-tune these open-weight models on a small pathogen corpus to construct a high-precision biological design tool in under 24 GPU hours. Similarly, Protein Language Models (PLMs) can be easily adapted to optimize peptide toxins or bypass natural immune barriers. Currently, the scientific community lacks a middle ground between full open-weight release (which exposes dangerous dual-use capabilities) and complete withholding (which stifles scientific research). There is an urgent need for a mechanism that guarantees open-weight models cannot be successfully repurposed for dangerous biosecurity threats. We propose SafeBio-Registry, an open-source verification and compliance platform (integrated as a Hugging Face Space and repository registry) designed to certify that hostedbiological models are safe and robustly protected against malicious exploitation. The platform will operate via a three-tiered pipeline: 1. Automated Safety Auditing & Verification: Genomic Language Models (GLMs): Automatically audited against biological benchmark suites (such as the Human Virome Understanding Evaluation, or HVUE) to guarantee they do not possess out-of-the-box pathogen design optimization or highly structured virological capabilities. Protein Language Models (PLMs): Screened for toxicity modeling and design capabilities to ensure they are not pretrained or easily prompted to generate dangerous bio-toxins or hazardous biochemical compounds. 2. Automated "SpecDef" Weight Locking Implementation: Drawing from our team's attached research, "Safeguarding open-weight genomic foundation models through weight locking," we will implement an automated pipeline that applies Spectral Deformation (SpecDef) to the write-side output projection matrices of the uploaded model's blocks. SpecDef inflates the top $k_{\text{lock}}$ singular values of targeted weight matrices by a factor $\alpha$, paired with a compensation matrix $C$ that preserves the exact forward pass at inference. This mathematically constructs an ill-conditioned loss landscape, forcing the stable learning rate threshold to drop precipitously ($\eta \lesssim 1/\alpha^2$). If a malicious actor attempts standard fine-tuning on a locked model to teach it pathogenic or toxic properties, the model will either actively suppress its capability (driving the Area Under the Receiver Operating Characteristic, or AUROC, below the pretrained baseline) or suffer immediate gradient explosion and divergence. An informed attacker attempting an advanced SVD-chain factorization bypass will be forced to pay a prohibitive computational cost making mass exploitation economically and technically impractical. 3. SafeBio Certification and Registry: Models passing the validation and locking pipeline will receive a cryptographic "SafeBio-Certified" badge. The platform will host the weight-locked version of the weights directly, creating a centralized ecosystem of secure-by-design biological models that researchers can safely download, deploy, and utilize.