Patronus AI finds ‘alarming’ safety gaps in leading AI systems

As generative AI technology continues to advance, concerns about safety and ethical risks have escalated. Patronus AI, a startup specializing in responsible AI deployment, unveiled SimpleSafetyTests, a diagnostic tool designed to identify critical safety vulnerabilities in large language models (LLMs) like ChatGPT.

The SimpleSafetyTests diagnostic tool comprises 100 handcrafted test prompts, addressing five high-priority harm areas, and has revealed significant safety variations across different language models. While certain models demonstrated flawless performance, others struggled with over 20% of test cases, raising concerns about their reliability in steering users away from harm.

Patronus AI emphasizes the importance of AI safety testing and mitigation services to ensure the responsible use of LLMs. The release of SimpleSafetyTests aligns with the increasing demand for ethical and legal oversight in AI deployment, with experts advocating for regulatory bodies to collaborate with industry players to produce safety analyses and evaluation reports.