Presentation: Leveraging Adversary Emulation for GenAI Red Teaming
| Source: InfoQ AI/ML
Tags: GenAI security, red teaming, MITRE ATLAS, AWS Bedrock, LLMjacking, Mitigant, data poisoning
Mitigant CTO Kennedy Torkura walks through practical GenAI red teaming for AWS Bedrock deployments, covering data poisoning, LLMjacking, and MITRE ATLAS-mapped attack techniques for security and cloud engineering teams new to securing AI workloads.
Details
In a 36-minute InfoQ conference presentation, Kennedy Torkura, CTO and co-founder of cloud security startup Mitigant, explains how engineering teams can apply traditional cloud security thinking to GenAI threat modeling. The core framework is MITRE ATLAS, an AI-specific attack taxonomy modeled on MITRE ATT&CK, which maps real-world AI attack cases to reusable threat patterns with documented remediation steps.\n\nKey threat categories covered include AI model tampering via supply chain attacks (drawn from MITRE ATLAS case studies), data poisoning of knowledge bases, and LLMjacking — where attackers exploit API access to run unauthorized inference at the victim's expense. Torkura frames this as Red Teaming 101, bridging traditional security engineering skills to AI-specific threat modeling.\n\nThe AWS Bedrock focus is practical: teams using managed LLM APIs need different red teaming approaches than those running self-hosted models. MITRE ATLAS provides a structured starting framework rather than ad hoc probing. Torkura also covers defensive guardrail implementation alongside offensive techniques, making this useful for both red team and blue team practitioners.