Research, benchmarks, and open-source tooling for LLM threat modeling.
Twenty-seven peer-reviewed papers, the PromptArmor open-source project, MITRE ATLAS coverage evaluations, and the commissioned Forrester Total Economic Impact study — everything a skeptical CISO needs to validate our claims before booking a demo.
LLM threat modeling research catalog.
The corpus covers prompt injection, model exfiltration, training-data leakage, and supply-chain attacks across OpenAI, Anthropic, Bedrock, Vertex, and self-hosted foundation model stacks. Each entry lists authors, venue, year, and a one-line abstract so you can scan relevance without leaving the page.
-
Coverage evaluation of LLM threat detection across the MITRE ATLAS adversary matrix.
Independent benchmark of detection depth across all 14 ATLAS tactics — Shaiex surfaces 4.8× more shadow LLM endpoints than the nearest competitor.
prompt injection exfiltration ATLAS Featured142 citations -
A taxonomy of indirect prompt-injection payloads against retrieval-augmented assistants.
Categorizes 1,184 adversarial payloads from in-the-wild RAG deployments and proposes the PU-12 classification still cited by OWASP LLM Top-10.
prompt injection RAG OWASP98 citations -
Training-data leakage in production foundation models: measurement and mitigation.
First systematic study of verbatim regurgitation across OpenAI, Anthropic, and open-weight models; introduces the RecallBench harness now maintained at Shaiex.
data leakage benchmark foundation models211 citations -
Model exfiltration via accelerator side channels in multi-tenant inference clusters.
Demonstrates weight recovery from co-tenant GPU memory pressure signals; informs the isolation guarantees in Shaiex Defender's data-plane.
model exfil side channel GPU76 citations
Showing 4 of 27. Full catalog with DOI links, BibTeX, and PDF access is available to verified enterprise evaluators.
Request full access →Why we publish.
When we started Shaiex in 2020, the security community had no shared vocabulary for LLM-specific attacks. Threat models written for traditional software did not transfer. Red-teamers were improvising. We thought the fastest way to fix that was to publish every technique we found — including the ones our own product defends against — and let the field argue about them in the open.
Six years later, the research program has produced twenty-seven peer-reviewed papers, the PromptArmor open-source project, and the LLM threat-detection track at USENIX Security. We are proud of the work, and prouder still that researchers outside Shaiex have built on top of it. PromptArmor alone has been forked into three commercial products and is referenced in the OWASP LLM Top-10.
What we're publishing next: a four-part series on agentic-tool exfiltration (Q2 2025), an updated MITRE ATLAS coverage matrix for multimodal models (Q3 2025), and the v2 PromptArmor release with native support for streaming token-level filtering. If you want the preprints before they go public, the best way in is to book a demo and tell us what you're trying to defend.
2024 MITRE ATLAS coverage evaluation.
An independent benchmark of detection depth across all 14 MITRE ATLAS adversary tactics. The study measured 612 enterprise rollouts across financial services, healthcare, and federal contracting — and Shaiex surfaced 4.8× more shadow LLM endpoints than the nearest competitor across initial-access, persistence, and exfiltration tactics.
- Peer-reviewed at: MITRE ATLAS working group, arXiv:2411.04821
- Methodology: black-box replay of 1,184 documented ATLAS techniques against live detection surfaces
- Coverage: OpenAI, Anthropic, Bedrock, Vertex, and self-hosted stacks
- Reproducibility: full evaluation harness open-sourced under MIT
PromptArmor: the open-source prompt-injection firewall.
MIT-licensed, maintained by the Shaiex research team, and used by security teams at companies we don't sell to. PromptArmor ships as a single-binary sidecar that wraps any LLM gateway — OpenAI, Anthropic, Bedrock, Vertex, or self-hosted — and blocks the OWASP LLM Top-10 at the prompt boundary with a 3 ms median latency overhead.
- 8,400+ GitHub stars, 312 contributors, MIT licensed
- 99.4% block rate across OWASP LLM Top-10 published benchmarks
- 3 ms median latency overhead — no extra hop in the hot path
- Drop-in for LangChain, LlamaIndex, OpenAI Functions, and MCP tool calls
"PromptArmor is the first prompt-injection control we'd actually let a customer run in production without our team in the loop."
— Staff Security Engineer, Fortune 50 insurer (independent quote, used with permission)