Adversarial lab
Research that changes the rule set.
We study attacks against language models, agents and the tools they call, then turn the findings into detections and policy. Research notes are published only after the mitigation exists.
Prompt injection
Direct and indirect payloads, encoding tricks, multi-turn escalation and the difference between a payload the model reads and an instruction it follows.
Agent abuse
Goal hijacking, tool confusion, confused-deputy calls through MCP servers, and the action policies that stop a compromised plan from becoming a transaction.
Evidence
What a decision record must contain to survive an incident review: input hash, controls, verdict, policy version, actor and action.
Disclosure
Report a bypass
If you find a novel injection, a jailbreak chain or an agent escalation path in this product, report it to security@intellirity.com. Include the affected surface, a minimal reproduction and whether customer data is involved.
We acknowledge good-faith reports, will not pursue legal action for research conducted within scope, and add validated findings to the detection policy.
Read the security modelField notes