Anthropic published a detailed cybersecurity safeguards document and a draft AI jailbreak severity framework on July 2, two days after globally redeploying Claude Fable 5 (Anthropic). The framework, co-developed with Glasswing partners including Amazon, Microsoft, and Google, proposes a five-band Cyber Jailbreak Severity (CJS) scale from CJS-0 (Informational) to CJS-4 (Critical) (AI Weekly). Each finding is scored on four axes - capability gain, breadth of impact, ease of weaponization, and independent discoverability - with summed scores mapping to bands on an exponential rather than linear scale, so each tier represents a materially higher risk than the preceding one (CyberSecurityNews). Fable 5’s existing cybersecurity classifiers sort requests into four categories rather than applying blanket refusals, a design Anthropic says is intended to preserve legitimate security research utility while blocking weaponization paths (CryptoBriefing). Anthropic framed the scale as a starting point for cross-industry standardization, explicitly scoping it to cybersecurity jailbreaks and excluding other misuse categories. Alongside the framework, Anthropic launched a HackerOne program that accepts submissions of potential cyber jailbreaks in Fable 5 under a published vulnerability disclosure process (HackerOne).
Claude models became generally available on Microsoft Azure Foundry on July 1, running on NVIDIA GB300 NVL72 systems with Quantum-X800 InfiniBand networking - the first deployment of Anthropic’s models on NVIDIA’s Blackwell Ultra generation of accelerators (Microsoft Azure Blog, NVIDIA Blog). The GB300 platform adds 50% more low-precision compute and doubles attention-layer throughput compared with the prior Blackwell generation, with Quantum-X800 InfiniBand designed for the multi-node agent workloads Microsoft positions as the primary enterprise use case (NVIDIA Blog). Claude Sonnet 5, Opus 4.8, and Haiku 4 are all available through Foundry Agent Service, which uses Claude as the reasoning core for multi-step planning and tool orchestration across enterprise systems (Neowin). Sonnet 5 carries introductory pricing of $2 per million input tokens and $10 per million output tokens through August 31, after which standard rates of $3 and $15 apply; prompt caching, extended thinking, and tool streaming are all supported in the Foundry deployment, addressing a procurement barrier for regulated-industry enterprise customers (TechTimes).