A new class of AI-specific malware called 'promptware' exploits the architectural flaw in LLMs where instructions and data are indistinguishable. Based on Bruce Schneier's research, the promptware kill chain mirrors traditional cyberattack stages: initial access via prompt injection, privilege escalation through jailbreaking, reconnaissance of connected tools and APIs, persistence via long-term memory stores like RAG databases, command-and-control through the LLM's internet access, lateral movement across interconnected agents, and final action on objective such as data theft or code execution. Defenses require a zero-trust approach — treating AI agents as untrusted execution environments, limiting tool access, constraining privileges, and designing systems to break the kill chain at each stage rather than relying on preventing initial access.

10m watch time
1.3K Impressions