AI threat research
Soteria Labs studies how attackers target LLMs, RAG systems and autonomous agents. What we learn goes straight into our detection models, our attack library and our customer guidance.
Research areas
Prompt Injection
Taxonomy of direct and indirect injection, and detection methods that hold up against paraphrasing and encoding.
Agentic Threats
Tool poisoning, goal hijacking, privilege escalation and multi-agent failure modes.
Data Extraction
System prompt leakage, membership inference and training data extraction.
Multilingual & Multimodal Attacks
Attacks that use low-resource languages, images, audio and documents to bypass filters.
Key findings
Indirect injection is now the dominant risk for enterprise AI
In enterprise deployments, the most dangerous instructions rarely come from the user. They come from content the model reads: emails, web pages, tickets and shared documents. Defenses have to inspect retrieved and tool-returned content, not only user prompts.
Agent permissions are the real blast radius
Whether a jailbreak causes harm depends far more on what the agent is allowed to do than on the model itself. Least-privilege tool authorization reduces impact even when detection fails.
Static guardrails decay quickly
Attack techniques evolve every week. Guardrail policies that aren't continuously re-tested lose effectiveness over time, which is why we pair runtime protection with continuous red teaming.
Multilingual gaps are widespread
Safety filters tuned mainly on English often miss the same attack phrased in another language or mixed-language text. Global enterprises need multilingual evaluation.
Responsible disclosure
When our research uncovers vulnerabilities in third-party models or tools, we follow coordinated disclosure with the affected vendors before publishing details.
Put the research to work
Our findings are built into Soteria RedTeam and Soteria Guard.