Anthropic has unveiled Constitutional Classifiers++, a major upgrade to its safety systems designed to prevent users from bypassing AI guardrails—a practice known as jailbreaking. While previous security versions significantly reduced harmful outputs, they often slowed down the model and accidentally blocked perfectly safe questions, creating a frustrating experience for users. This new iteration solves those friction points by using a cascade architecture, a two-stage screening process that acts like an intelligent security checkpoint to filter traffic efficiently. The first stage uses linear probes to monitor the model’s internal neural activations. Essentially, the system peeks at Claude’s gut intuition to see if it detects a harmful pattern or suspicious signal before the AI even finishes formulating its final response. This method is incredibly resource-efficient, adding only about 1% to total compute costs. If the probe flags a suspicious exchange, it escalates the query to a more sophisticated exchange classifier that analyzes both the user's prompt and the model's answer in tandem. This context-aware approach is crucial for stopping obfuscation attacks, where users hide dangerous requests behind riddles, metaphors, or code words like using "food flavorings" as a substitute for toxic chemicals. By looking at the full conversation history, the classifier can spot the hidden intent that a single-sided check would likely miss. The result is a system that is both more secure and significantly less intrusive, boasting an 87% reduction in false-positive refusals while remaining resilient against universal jailbreaks.

Anthropic has unveiled Constitutional Classifiers++, a major upgrade to its safety systems designed to prevent users from bypassing AI guardrails—a practice known as jailbreaking. While previous security versions significantly reduced harmful outputs, they often slowed down the model and accidentally blocked perfectly safe questions, creating a frustrating experience for users. This new iteration solves those friction points by using a cascade architecture, a two-stage screening process that acts like an intelligent security checkpoint to filter traffic efficiently. The first stage uses linear probes to monitor the model’s internal neural activations. Essentially, the system peeks at Claude’s gut intuition to see if it detects a harmful pattern or suspicious signal before the AI even finishes formulating its final response. This method is incredibly resource-efficient, adding only about 1% to total compute costs. If the probe flags a suspicious exchange, it escalates the query to a more sophisticated exchange classifier that analyzes both the user's prompt and the model's answer in tandem. This context-aware approach is crucial for stopping obfuscation attacks, where users hide dangerous requests behind riddles, metaphors, or code words like using "food flavorings" as a substitute for toxic chemicals. By looking at the full conversation history, the classifier can spot the hidden intent that a single-sided check would likely miss. The result is a system that is both more secure and significantly less intrusive, boasting an 87% reduction in false-positive refusals while remaining resilient against universal jailbreaks.