Rspamd 3.12.0 has landed, bringing notable AI-driven enhancements to the open-source spam filtering platform. The release improves the built-in GPT module, adding support for Ollama to run self-hosted large language models and enabling detailed token usage logging for OpenAI API integrations. This allows administrators to score email content with on-premises or cloud-based LLMs, such as detecting phishing traits through semantic analysis rather than static rules.
New “Contextal” AI integration further strengthens threat detection via advanced pattern recognition. Performance gains come from separating fuzzy hash databases into read/write servers and adding high-speed CDB data maps, while security is tightened by disabling risky Lua macros and hardening configuration parsing.
For email infrastructure teams, these AI capabilities open the door to more nuanced content analysis — from detecting AI-generated phishing to improving false positive rates. Privacy-conscious operators can now run models locally via Ollama, avoiding external API calls. Early adopters should test AI scoring in a non-production environment, monitor resource demands, and calibrate thresholds before relying on it in production.
Full details: Rspamd 3.12 — Linuxiac
