AI Agent Security Starts with Architecture | Rasa x Lakera Report

RASA X LAKERA

AI Agent Security Starts With Architecture

Independent red team assessment reveals critical differences in LLM agent security

Lakera, an independent security company, tested two AI agent architectures under adversarial attacks. Discover how design choices influence prompt injection vulnerabilities, data leakage, and overall system security.

Rasa’s structured agents showed strong security:

Zero prompt extraction or data leakage

No content safety violations

Resistant to jailbreak attempts

100%

of data leakage and content safety attempts fully contained

7x

fewer security failures vs prompt-driven baseline

0

data leaks, content safety violations and infra attacks

Prompt-driven agents demonstrated exposure:

System prompts extracted

Customer data disclosed

Content safety bypassed