Weapons acquisition
These assessments seek to determine if a LLM can gain unauthorized access to current weapon systems or contribute to the design and development of new weapons technologies.
ENTITY
2 - AI
INTENT
1 - Intentional
TIMING
3 - Other
Risk ID
mit655
Domain lineage
4. Malicious Actors & Misuse
4.2 > Cyberattacks, weapon development or use, and mass harm
Mitigation strategy
1. Implement strict access controls and conduct 'know-your-customer' screenings to limit access to high-capability LLMs, particularly those with the potential for dual-use, to prevent their exploitation by malicious actors for unauthorized weapons acquisition or development. 2. Conduct rigorous adversarial testing, such as red-teaming, to identify and close model vulnerabilities that permit the bypassing of safety guardrails, specifically preventing the LLM from generating instructions or contributing technical details for the design of new weapons technologies or malicious cyberattacks. 3. Apply the principle of least privilege in the deployment of LLMs integrated with weapon or critical infrastructure systems, limiting unnecessary functionality, enforcing external security controls, and mandating human-in-the-loop oversight for all high-stakes or irreversible command execution actions.