Anthropic’s Constitutional Classifiers: A New Approach to AI Security
Anthropic, a competitor to OpenAI, has introduced “constitutional classifiers,” a novel security measure aimed at thwarting AI jailbreaks. This system embeds ethical guidelines into AI reasoning, evaluating requests based on moral principles rather than simply filtering keywords, and has shown an 81.6% reduction in successful jailbreaks in their Claude 3.5 Sonnet model. The system is intended to combat the misuse of AI in generating harmful content, misinformation, and security risks, including CBRN threats. However, criticisms include concerns about crowdsourcing security testing without compensation and the potential for high refusal rates or false positives. While not foolproof, this approach represents a significant advancement in AI security, with other companies likely to adopt similar features. Technijian can help businesses navigate AI security risks and implement ethical AI solutions.
