Selected GenAI security resources (November 2026)
Top GenAI security resources — November 2026
Adversa AI released a November 2026 GenAI security resource compilation, organizing 24 resources by theme, with defenses the most numerous at 9. The compilation notes that most of this month's specific failures occurred in loading and serving code beneath the models; for example, a large request to GLM-5.3 on Cloudflare Workers AI could return other tenants' inference content, with the flaw traced to upstream SGLang, and Cloudflare has fixed it; mining and backdoor activity across 31 model repositories on Hugging Face reappeared within days of being taken down. On the attack side, a tree search method bypassed all 6 multi-scanner guardrail systems across 120 scenarios. The checklist at the end recommends upgrading vLLM to 0.32.0, upgrading unsloth_zoo to 2026.8.14 or higher, checking code and CI for trust_remote_code=True, and retesting guardrails separately by injection channel.
- Context-Aware Prompt Injection Detection in RAG Pipelines with a Small Language Model: nokast-secureRAG
- CorrectGuard proposes a view-free correctness estimation framework for black-box safety guardrails
- MOMAT: A multi-atlas mixture framework for low-power jailbreak defense of quantized LLMs
- Persona Guardrail: A production-grade defense framework for agent systems
- Study: Classification suffixes improve out-of-distribution malicious input detection with activation probes