Anthropic捨棄關鍵AI安全承諾

🔑 增強重點摘要

• •Anthropic's new Responsible Scaling Policy (RSP v3.0) introduces a 'Responsible Scaling Policy' framework separate from industry-wide recommendations, modeled after US government biosafety level (BSL) standards, enabling more granular risk assessment across different capability thresholds[1][4].
• •The policy shift was driven by an 'anti-regulatory political climate' and lack of federal AI governance progress; Anthropic's leadership concluded that unilateral safety pauses would disadvantage responsible developers while weaker competitors set the pace for the industry[1][2].
• •Anthropic is implementing new technical safeguards including fully automated attack investigation systems to detect coordinated misuse patterns, confidential compute adoption across model R&D lifecycles, and AI-assisted security tooling for vulnerability discovery and anomaly detection, with initial projects due by April 1, 2026[3][4].
• •The revised policy mandates external third-party expert review of Risk Reports under certain circumstances, with reviewers receiving unredacted or minimally-redacted access to Anthropic's safety analysis and decision-making processes[4].

🛠️ 技術深入

• •ASL-3 Security Standard and Deployment Standard: Enhanced safeguards triggered by specific Capability Thresholds, including CBRN (Chemical, Biological, Radiological, Nuclear) development capabilities and AI R&D automation milestones[4][5].
• •Input and output classifiers: Anthropic developed sophisticated methods to block concerning content, particularly for ASL-3 deployment standards targeting chemical and biological weapons risks from threat actors with modest resources[4].
• •Capability Thresholds tracked include: (1) ability to fully automate entry-level AI research work, (2) ability to cause dramatic acceleration in effective scaling rates, and (3) capabilities that could uplift moderately resourced state CBRN programs[5].
• •Planned safeguards include centralized records of critical AI development activities analyzed by AI systems for insider threats and security vulnerabilities, plus a 'regulatory ladder' policy framework for government guidance[4].
• •Confidential compute feasibility analysis and continuous personnel security vetting programs for high-risk roles with defined screening criteria and monitoring requirements are under development[3].

🔮 前景展望AI analysis grounded in cited sources

⏳ 時間線