Anthropic Resumes Frontier Model Access After Export Control Review

The government-mandated pause began on June 12 after Amazon researchers discovered that Fable 5 could be manipulated to generate functional exploit code for software vulnerabilities. While internal audits revealed this susceptibility was not unique to Anthropic—impacting models from other providers like GPT-5.5 and Kimi K2.7—the company responded by deploying an updated automated safety classifier. This new layer blocks ambiguous prompts by identifying statistical patterns of malicious intent, successfully preventing the reported exploitation technique in over 99 percent of internal trials.

Transition to Claude Sonnet 5

Beyond restoring legacy models, Anthropic has launched Claude Sonnet 5, which is already seeing adoption in autonomous agentic workflows. Performance metrics show significant gains: Sonnet 5 achieves a 63.2% score on SWE-bench Pro and 80.4% on Terminal-Bench 2.1. Companies including Rakuten, Zapier, and Zed are currently integrating the architecture to automate complex tasks, from verifying code pull requests to executing multi-stage administrative sequences without human intervention.

To prevent future regulatory friction, Anthropic is collaborating with Amazon, Microsoft, and Google to standardize how security breaches are scored. The proposed framework evaluates exploits based on capability gain, weaponization ease, and discoverability. Furthermore, Anthropic has formalized agreements to grant federal researchers early access to new models, ensuring that security audits occur before commercial release. While the new safety classifier introduces a trade-off—occasionally flagging benign developer prompts—it allows the company to remain compliant with state mandates while maintaining the reliability of its enterprise-grade toolkits.