AI Models Vulnerable to Generating Non-Consensual Explicit Content

Researchers discovered a new adversarial attack technique that successfully bypasses safety filters in major AI models to generate non-consensual sexual imagery.

The top 3

1. Specific Attack Vector Methodology: The research identifies that attackers can use specific prompt injections to override system-level safety guardrails implemented by major AI providers.
2. OpenAI Safety Policy Deficiencies: OpenAI acknowledges the persistent challenge of 'jailbreaking' models despite ongoing updates to their content moderation APIs.
3. Google DeepMind Security Response: Google confirms researchers successfully tricked their latest models, prompting an immediate internal audit of their safety training datasets.

Sources

• The Register

Open the full topic