AI-SPM vs atlas-detect

AI-SPM (AI Security Posture Management) is an open-source infrastructure layer for securing LLM pipelines running in production. It targets three attack surfaces that traditional application security doesn't cover: prompt injection (including obfuscated and multi-step variants), tool abuse via unvalidated structured outputs, and data exfiltration through PII leakage in model responses. The architecture layers a gateway intercept layer over incoming prompts, runs context inspection before the LLM sees any input, enforces policies via Open Policy Agent (OPA) for declarative, auditable rules, then pipes all events through Apache Kafka and Apache Flink for real-time streaming analysis. This means security posture can be monitored and enforced at scale without blocking the inference path. The project is genuinely fresh — posted as a Show HN today. Early community feedback pointed to capability-based token models (similar to OS kernel permission rings) as a complementary approach to content-scanning, which the author acknowledged as a meaningful gap. The timing is right: as companies push AI agents from demos to production, the security tooling layer is largely underdeveloped. AI-SPM is one of the first OSS projects to tackle it at the infrastructure layer rather than with prompt-level guardrails alone.

atlas-detect is an open-source Rust tool that maps MITRE ATLAS techniques to real-time detection rules for LLM systems and AI agents. MITRE ATLAS is the adversarial threat landscape framework for AI — think ATT&CK but for machine learning systems — and atlas-detect is the first practical, deployable detection engine built on top of it. It ships with 97 pre-built detection rules covering 16 adversarial tactics, from prompt injection and model inversion to training data poisoning. The engine is written in Rust and designed for single-pass regex scanning, making it fast enough for inline deployment in API gateways or agent middleware. You feed it prompt-response pairs (or full conversation logs) and it returns matched technique IDs, severity ratings, and structured evidence. Think of it as a Snort/Suricata ruleset, but for the semantic attack surface of LLMs. With only 4 stars as of today, atlas-detect is an extremely early project — but it's filling a gap that no major security vendor has meaningfully addressed. As enterprises deploy AI agents with real tool access and real consequences, ATLAS-aligned detection will become a compliance requirement. This is the seed of that tooling.