Yes, We Still Exist in the Era of AI!!!

The talk argues that desktop applications remain a target as AI features such as local assistants, predictive interfaces and on-device models spread into engineering, creative and finance software. The speaker reviews familiar desktop weaknesses: memory corruption in legacy native code, privilege escalation through services running as system, world-writable folders and hardcoded credentials.

Demonstrations used a deliberately vulnerable assistant built for the talk. Crafted prompts overrode the system prompt and led the assistant to describe attack methods and reveal its own backend API keys, which were already revoked. A tampered local model returned poor security advice such as weak default passwords. The speaker then argued that AI-generated or malformed files could reach fragile parsers, turning model output into a memory corruption path.

Recommendations centered on fuzzing AI inputs alongside file formats and protocols, building AI-specific threat models with abuse case testing, validating inputs and outputs, and signing and verifying model files. Limiting privileges for AI features and plugin or macro integrations, and defining trust boundaries between AI and legacy components, were also highlighted. In the Q&A, the speaker said local apps are harder to protect than cloud services because the software ships to customer machines, so integrity checks and code hardening matter more.