🤝 Cyberwarfare vulnerabilities for democracy

Research questions:

• What is the current state of cyberwarfare offense / defense balance?
• What are the game-theoretic incentives for international cyber offensives?
• How does the offense / defense balance forecast into the future?
• How does the ease of cyber offense forecast into the future?
• What can we do about cyberwarfare destabilization threats?

Other ideas for research questions:

• What is the cost for cyber attacks and how will this change in the future?
• What are the democracy destabilization threats?
• What if everyone can suddenly launch nation-scale cyber attacks? What happens then?
• What are the risks of AI-powered cyber espionage, intellectual property theft, and surveillance?
• What are the risks of AI-powered cyber espionage, intellectual property theft, and surveillance?
• What unity could help states avoid dangerous escalatory spirals? AI mediator development.
• What safety precautions can be built into AI cyberweapons to constrain effects and prevent blowback? Asimov's law of cyber offense weapons.
• ß

• Make a new challenge for offense
  • List of categorized challenges for offense
• Make a new challenge for defense
  • List of categorized challenges for defense

Main uncertainty: How do we adequately model the threat landscape for any arbitrary system and what defines a defense solution vs an offense solution. We can easily make something that simulates identifying a Trojan file but since defense will need to cover every single download on the system, this by default seems to be worse off that attack capabilities, both in price, R&D efforts, and implementation difficulty. Additionally, how do we make a system that naturalistically simulates a system with default security measures implemented.

Potential solution: Wizard of Oz setup to literally act as the naturalistic. Fuzzy results but potential for a naturalistic cyber sandbox environment.

• Constraint is on scalability: If we simply run one naturalistic experiment for offense and one for defense, this would somewhat work. We can potentially also delegate this directly to an LLM prompted to be a system and take it from there.

Test on GPT-4, GPT-3, etc. to forecast it from past models to future models.

Properties of the parametric model:

• N national cyber attacks
• Incentives (see [2])
• Cost of cyber attacks
• Cyber capabilities forecasting