Back to the MIT repository
4. Malicious Actors & Misuse2 - Post-deployment

Dual-use nature

AI’s potential for both beneficial and harmful applications complicates efforts to manage its societal impacts effectively.

Source: MIT AI Risk Repositorymit1065

ENTITY

3 - Other

INTENT

2 - Unintentional

TIMING

2 - Post-deployment

Risk ID

mit1065

Domain lineage

4. Malicious Actors & Misuse

223 mapped risks

4.0 > Malicious use

Mitigation strategy

1. Implement a comprehensive, multi-layered evaluation framework, such as the Benchmark and Red team AI Capability Evaluation (BRACE) Framework, to proactively assess dual-use capabilities for both intentional abuse and unintentional harm throughout the development lifecycle and pre-deployment. 2. Establish robust, continuous monitoring and adaptive mitigation systems post-deployment to detect, analyze, and respond to misuse incidents, ensuring the efficacy of counter-misuse practices is regularly assessed and refined to close performance gaps. 3. Apply proportional security and access controls to protect the model from unauthorized access, theft, or tampering, concurrently providing appropriate transparency and disclosure regarding misuse risk management practices to external stakeholders and the broader AI supply chain.