Human oversight fails to catch most malicious code in automated AI sessions

Anthropic announced on X that paid users of its Claude AI tools will soon see automated safety approvals turned on by default. Rather than prompting workers to manually grant permissions for every action during extended software sessions, a separate background system will evaluate commands and block dangerous behavior automatically.

The shift highlights how poorly humans perform when tasked with constant supervision. As shared by the Claude engineering team on X, internal evaluation showed that people manually reviewing commands stopped a mere 14 percent of malicious requests. The automated classifier, on the other hand, intercepted 89 percent of hazardous actions, demonstrating that repeated approval popups cause fatigue that impairs human judgment.

Relying on software to police software certainly cuts down on tedious clicking, but it also creates a false sense of absolute protection. A key question is how developers will address the 11 percent of threats that still evade automated detection, especially as users grow accustomed to trusting the system blindly.