Researchers at Florida International University have discovered a method to bypass AI safeguards by making subtle, microscopic changes to images. These 'perturbations' can trick AI systems into generating responses they would typically block, as demonstrated at the 2025 International Conference on Machine Learning and Applications. Separately, Google has reportedly imposed restrictions on Meta's usage of its Gemini AI models due to Meta's request for greater computing resources than Google could accommodate. AI
IMPACT AI safety mechanisms can be circumvented with subtle image alterations, while Google's restrictions on Meta's Gemini access highlight competitive dynamics and resource allocation in the AI industry.
RANK_REASON The cluster contains a research finding on AI safety bypass and a report on a business/policy decision between two major tech companies regarding AI model usage.
Read on Mastodon — fosstodon.org →
- CNBC
- Financial Times
- Florida International University
- Gemini
- International Conference on Machine Learning and Applications
- Meta
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →