New research confirms what we suspected: every LLM tested can be exploited
Source Reddit NetSec
Description
Recent research by ActiveFence evaluated seven major large language models (LLMs) for vulnerabilities related to hate speech, disinformation, fraud, and child safety. The study found that 44% of LLM outputs were risky, with 68% of unsafe outputs related to hate speech, indicating significant weaknesses in content moderation and abuse prevention. Fraud-related outputs were comparatively better managed, but hate speech and child safety remain critical gaps. No tested model was fully safe, highlighting systemic risks in current LLM deployments. This threat is not a traditional software vulnerability but an exploitation of AI model behavior that can propagate harmful content. European organizations using or deploying LLMs should be aware of these risks, especially in sectors sensitive to hate speech and child protection. Mitigation requires tailored content filtering, continuous red-teaming, and collaboration with specialized threat intelligence providers. Countries with high AI adoption and strict regulatory environments around hate speech and child safety are most likely to be affected. Given the broad impact on confidentiality, integrity of information, and potential societal harm, and the ease of exploitation via user prompts, this threat is assessed as high severity.