Exploiting Multi Agent Systems: How Prompt Injection Turns Collaboration into Compromise

Placeholder intro | Conference briefing

Presenter: Jeremy Richards
Team: AI Red Team, ServiceNow

Intro

Large language model agents no longer operate in isolation. They collaborate, delegate, and execute tool calls across orchestration layers, which creates a larger security surface than single-agent prompt injection alone.

This briefing introduces a red-team attack path for multi-agent systems: harvesting hidden system prompts, escalating through mirrored pattern injections, subverting planner behavior, and influencing downstream tool use. The session focuses on both direct prompt injection and second-hand (indirect) injections that move across agent boundaries and compound into mission-level compromise.

Session details

• Date and time: Thursday, October 2, 3:15 PM to 4:00 PM
• Room: 801B
• Format: 45-minute briefing
• Tracks: AI, ML, and Data Science; Application Security

Scope and deliverables

• Attack chain walkthroughs for direct and indirect multi-agent injection
• Practical defender controls with trade-offs in engineering effort and efficacy
• Orchestrator policy patterns for reducing blast radius
• Observability patterns that support security telemetry while respecting privacy

Status

This page is a placeholder intro for the published session. A longer technical write-up and supporting materials will be added in a future revision.