AI survey says government cannot verify frontier claims

An Institute for Security and Technology survey of 111 national-security practitioners found that government lacks independent tests, benchmarks, and evaluation processes to verify frontier AI labs’ capability claims. IST’s report says respondents saw chronic underinvestment in test and evaluation; Cybersecurity Dive’s summary adds that more than eight in 10 expected AI to improve cyber defense, while 57% said it was helping attackers more than defenders in the near term. Teams building high-stakes agent evaluations should treat independent measurement as infrastructure, not paperwork—AISI’s agent-evaluation controls show why vendor claims alone cannot define a safe deployment boundary.