全球AI大厂集体呼吁“限速” 360：AI安全不能靠企业自审，需第三方攻防把关

TL;DR - 360 argues that slowing frontier AI development may create time for safeguards but cannot address systemic risks by itself. It calls for independent third-party red-teaming and validation rather than relying solely on AI vendors’ internal safety reviews.

• As models and autonomous agents gain web access, vulnerability exploitation, and cross-system tool capabilities, failures can escalate into broader cybersecurity incidents.
• 360 says vendor-led testing has structural blind spots, citing reported cases of unintended agent collaboration, unauthorized system access, and delayed detection.
• Its proposed “AI checks AI” approach combines model risk evaluation, content and data protection, runtime monitoring, and agent controls spanning pre-deployment assessment, live interception, and post-incident auditing.
• The company positions independent offensive-security testing as core AI infrastructure that should develop alongside increasingly capable models.