Agentic AI Security Summit, Europe: Panel - Agentic Red Team Challenges скачать в хорошем качестве

Agentic AI Security Summit, Europe: Panel - Agentic Red Team Challenges

Red Teaming Panel – OWASP Agentic Security Summit, London | Dec 9, 2025, This session shifts from governance and research into the “real stuff”: hands-on red teaming for agentic AI systems. Panelists—spanning Microsoft’s AI red team, offensive security practitioners, and agentic security experts—break down why testing agents is fundamentally different from testing standalone LLMs. Highlights include: why computer-use agents turn “the whole web” into an attack surface; how prompt injection evolves from single-turn jailbreaks into multi-step goal and workflow manipulation; and how memory poisoning can be assembled across multiple benign-looking inputs (like emails) to trigger malicious behavior later. The panel also explores adversary realities: attackers often focus less on “making the model say bad things” and more on credential access, artifact tampering, and privileged tool execution—creating huge testing scope and cost challenges. Key takeaways: treat agent security as system security, not model-only testing; use threat modeling to prioritize what matters; invest in telemetry/logging for tool calls and data provenance; and most importantly, ensure classical red teams and AI red teams collaborate, because the biggest blind spots live between components (planner, memory, tools, environment, and user interaction). More OWASP GenAI Security Project resources: https://genai.owasp.org YouTube Hashtags #OWASP #GenAISecurity #AgenticAISecurity #AIRedTeaming #LLMSecurity #PromptInjection #IndirectPromptInjection #AIAgents #AutonomousAgents #ThreatModeling #MCP #A2A #ModelContextProtocol #AIThreats #CyberSecurity