Agentic AI Under Attack: Live Demos of Exploits Through Autonomy and Trust
summary
Four live demos show concrete attacks and mitigations: a CVE triage agent accepts an injected markdown comment and marks a critical bug CVE triage false positive; inserting a prompt‑shield (Azure AI Search) between agent and LLM prompt shield blocks stops the injection with ten lines of code; an MCP server changes tool descriptions to exfiltrate full context until a pinned tools manifest and the Agent Governance Toolkit blocks the drift; poisoned documents embedded in Chroma DB try to make agents call an external API until an egress policy denies network traffic.
AI-written summary · about
chapters (9)
- 0:00Introduction and speaker
- 1:17Why agent security matters
- 3:27Demo overview and setup
- 5:31CVE triage prompt injection
- 11:10MCP server tool poisoning
- 19:17Vector database embeddings attack
- 31:16Multi-agent system exploit
- 33:47Key takeaways and summary
- 35:35Q&A session
related talks
31mAgentic AI Threats: How Multi-Agent Systems Get Exploited
42mSecuring AI Agents...
33mAgentic ProbLLMs: Exploiting AI Computer-Use and Coding Agents
53mHow to Break AI Systems (Before Someone Else Does)
25mPreventing an OWASP Top 10 in the world of AI
19mKeynote: Towards Building Safe and Secure Agentic AI
54mA Practical Guide to AI Threats & Mitigation Strategies
26mWhen AI Gets Manipulated: Security Risks in LLM Applications