Five frontier AI labs - Anthropic, OpenAI, Google, Microsoft, and Amazon - agreed on July 3 to adopt a shared Cyber Jailbreak Severity (CJS) scale by August 1, establishing the first cross-industry standard for classifying AI model jailbreaks by harm potential (TechTimes). The five-tier scale (CJS-0 to CJS-4) scores each incident on capability gain, breadth, ease of weaponization, and discoverability, with each tier representing exponentially more real-world risk than the level below it (Anthropic). The framework grew out of a 19-day U.S. export control order in June that forced Anthropic to pull Claude Fable 5 and Mythos 5 offline after Amazon researchers found a jailbreak enabling Fable 5 to flag software vulnerabilities and produce exploit-demonstration code (CNBC). All five labs also agreed to provide the NSA and CISA up to 30 days of pre-release access to covered frontier models before any broader partner rollout, under President Trump’s June 2 executive order (The White House).
Claude Fable 5 resumed global service on July 1 (Decrypt), but independent benchmarking has since found the new safety classifier causes significant regressions on security-adjacent coding tasks. BridgeMind reported Fable 5’s BridgeBench debugging score fell from 86.2 to 25.9 and refactoring from 73.6 to 38.4 after redeployment, with the drop traced to classifier routing rather than model degradation: only 3 of 12 TypeScript tasks in the test run reached Fable 5; the remaining 9 were silently rerouted to Opus 4.8, which scored zero by default (TechTimes). Prompts containing terms such as “vulnerability,” “exploit,” or “fix” are most likely to trigger the reroute. Anthropic confirmed the classifier over-flags benign requests and said it will refine the trigger rate without committing to a timeline.
Also on July 2, Anthropic launched the Claude Apps Gateway, a self-hosted control plane for enterprise Claude Code deployments on Amazon Bedrock and Google Cloud (Anthropic). The gateway runs as a stateless container backed by PostgreSQL and provides corporate SSO, centrally enforced policy controls, role-based model access, per-user cost attribution, and spend caps at the organization, group, or user level (DevOps.com). Inference routes to the Claude API, Amazon Bedrock, or Google Cloud with optional failover between providers, and usage data is not forwarded to Anthropic unless the operator configures the Claude API as an upstream.
xAI’s Grok 4.5 entered private beta at SpaceX and Tesla in late June, built on a 1.5-trillion-parameter V9 foundation model with supplemental training on Cursor IDE coding data - roughly three times the size of the 500-billion-parameter v8-small model currently handling production workloads on X (TechTimes). Elon Musk stated early evaluations place performance near or above Anthropic’s Opus tier, but no third party has accessed the model and xAI has not submitted it to any public benchmark suite, leaving performance claims unverified.