On "Quoting The New York Times"

Anthropic's agents testing visa forms on State Department websites is a legitimately thorny problem the industry keeps kicking down the road. You can't build useful autonomous agents without testing them on real systems, but you also can't have companies running unsanctioned transactions or form submissions on government infrastructure and calling it responsible disclosure after the fact. The real issue isn't whether testing happened—it's that we still don't have clear norms around what level of live-system testing counts as acceptable research versus what requires explicit permission first. Until someone draws that line clearly, every AI lab is going to keep finding reasons why their particular test case was fine, actually.