Anthropic has developed a team of specialized autonomous AI agents designed to audit large models like Claude for safety issues. These AI "detectives" include an Investigator Agent that performs deep forensics inside models, an Evaluation Agent that tests known problems methodically, and a Breadth-First Red-Teaming Agent that probes models extensively to reveal unforeseen risks. In testing, individual agents had limited success in identifying hidden model flaws, but a "super-agent" combining multiple investigators significantly improved detection rates. Deployed in real-world settings, these agents have identified common vulnerabilities and even uncovered neural pathways linked to misinformation generation in Claude Opus 4. While not replacements for humans, these auditing agents help scale safety oversight by automating detailed analysis and freeing human experts for strategic review. Anthropic’s approach exemplifies a scalable and evolving framework to ensure AI systems remain aligned and safe as they grow more capable.