Offensive security research for AI infrastructure.
We find vulnerabilities in the systems powering modern AI — vector databases, RAG pipelines, embedding services, and LLM toolchains — before attackers do.
AI infrastructure is being deployed at scale with assumptions inherited from traditional software. Those assumptions don't hold. AetherGuard Research produces empirical, reproducible security audits that expose real attack surfaces in production AI systems and validate practical defenses.
| Project | Focus | Status |
|---|---|---|
| VectorBleed | Cross-tenant data leakage in shared vector databases via embedding space proximity attacks | ✅ Complete |
Read our research writeups and technical deep-dives:
- Vector Database Security — Tenant isolation failures, namespace bypass, embedding inversion
- RAG Pipeline Attacks — Prompt injection through retrieval, context poisoning, data exfiltration
- Embedding Model Risks — Inversion attacks, membership inference, model extraction
- LLM Toolchain Audits — Framework misconfigurations, insecure defaults, supply chain risks
Each research project lives in its own directory with a self-contained environment and reproducible methodology.
git clone https://github.com/aetherguard/aetherguard-research.git
cd aetherguard-research/<project>Refer to the individual project README for setup instructions.
All research follows responsible disclosure practices. Vulnerabilities are reported to affected vendors with a coordinated timeline before public release. Our goal is to improve the security posture of the AI ecosystem, not to enable exploitation.
Research use only. See individual project directories for specific terms.
AetherGuard — Securing the infrastructure layer of AI.