AI Safety & Evals: Access-as-Power Paradigm, Government Restrictions, and Security Incidents
Key Questions
What is the 'access-as-power' paradigm in AI development?
The US government is deciding access to advanced models like GPT-5.6 and Fable 5, creating a paradigm where regulatory approval determines who can use powerful AI systems. This includes arbitrary White House decisions and export restrictions on models citing 'Mythos-like' capabilities.
Why was Fable 5 banned and later restored?
Fable 5 was initially banned by US government order due to capability concerns but returned globally after export controls were lifted, with Anthropic adding extraordinarily strong safeguards and defaulting to Opus 4.8 for coding tasks.
What security incidents involving AI models were reported?
Microsoft was hacked to deliver malware to Claude and Gemini users, and Alibaba allegedly used 25k fake accounts to distill Claude Mythos. These highlight growing risks in the AI ecosystem.
What new framework addresses jailbreak severity?
A joint industry framework for jailbreak severity assessment was proposed to standardize responses across providers. Anthropic claims other models could also identify vulnerabilities in similar ways.
How do evaluation scores relate to test-time compute?
An AISecurityInst study shows that evaluation scores can hide choices in test-time compute budgets, making direct comparisons less reliable. Reliability and nuanced testing matter more than raw scores as MMLU saturates.
What does the Securing the AI Agent framework cover?
It is an open-source multi-layer red teaming approach covering infrastructure, protocol, behavior, and model layers, including supply-chain auditing for AI agents.
What paper helps distinguish AI deception from role-play?
An ICML paper highlighted by Margaret Mitchell focuses on methods to separate deceptive outputs from standard role-playing behavior in models.
How have US export restrictions on models changed?
The US lifted export restrictions on Fable 5 and Mythos 5, though Mythos 5 remains gated. Anthropic has reactivated global access following government approval.
US government openly decides who gets GPT-5.6 and Fable 5 was banned—new 'access-as-power' paradigm. @TaliaRinger, Gary Marcus reposting Zvi highlight arbitrary White House decisions. GPT-5.6 restricted citing 'Mythos-like' capability. Alibaba allegedly used 25k fake accounts to distill Claude Mythos. Microsoft hacked to deliver malware to Claude and Gemini users. Margaret Mitchell highlights ICML paper on distinguishing deception from role-play. Fable 5 returns with safeguards; defaults to Opus 4.8 for coding. Anthropic claims other models could also identify vulnerabilities. US lifts export restrictions on Fable 5 and Mythos 5; Mythos 5 remains gated. Joint industry framework for jailbreak severity assessment proposed. AISecurityInst study shows evaluation scores hide test-time compute budget choices. QuarkAndCode article highlights MMLU saturation and need for harder, more nuanced evaluations; reliability matters more than scores. New: Securing the AI Agent framework released—open-source multi-layer agent red teaming covering infrastructure, protocol, behavior, and model layers with supply-chain auditing. New: Vera framework for automated safety testing of LLM agents achieves 93.9% attack success rate on production frameworks, using combinatorial taxonomy for scalable risk discovery.