5 stories tagged Auditing
Researchers from UC Berkeley and Stanford published a method to audit black-box LLM agents using a low-cost, open-weight surrogate model. The surrogate reads the same context and proposed action as the primary agent and provides a reliable error signal from its log-probabilities, catching mistakes before they execute.
A new study suggests that to properly audit AI agents, we must also examine the data they use to make decisions. This could change how we ensure AI systems are fair and reliable.
Researchers have developed a new method called DECOR to detect subtle deception in AI responses. This tool helps identify exactly how and where AI models distort facts, making it easier to hold them accountable.
Researchers propose a unified graph representation to track AI agents' decisions, making it easier to audit their actions. This could help ensure AI systems behave as intended and follow security protocols.
Researchers introduce Springdrift, a persistent runtime for LLM agents that prioritizes auditing. This could revolutionize how we track and verify AI agent actions.