RQ-20260702-001-agent-security
How do tool-using AI agents change the threat model compared with chat-only language model deployments?
Why It Matters
Agent systems can read, write, call tools, browse, execute code, and maintain memory. These capabilities expand both impact and attack surface.
Evidence Plan
- Survey prompt injection and tool abuse papers.
- Collect incident reports involving agentic workflows.
- Compare mitigations across sandboxing, permissioning, monitoring, and memory design.