4 stories this week that change your decisions (Jun 29-Jul 5, 2026)
TL;DR Anthropic redeployed Claude Fable 5 with a new safety classifier and lower blocking thresholds, after the US government forced it to revoke access when Amazon bypassed the model's safety filters to make it find vulnerabilities and write an exploit. Separately, Meta is standing up a cloud business it calls Meta Compute to sell its excess AI compute against AWS, Azure, and Google Cloud, and it acqui-hired the founders of Virtue AI, an enterprise AI-security startup, into Meta Superintelligence Labs. And a security review of five agent protocols found 30 additional failures that emerge only when the protocols are composed, with no party owning the full attack path.
1. Claude Fable 5 is back with tighter cybersecurity blocks
Anthropic redeployed Fable 5 with tighter cybersecurity blocks and a proposed industry framework for rating how bad AI jailbreaks are.
2. Meta is entering the enterprise security market
A cloud business to rival AWS, Azure, and Google Cloud, and a top AI-security team pulled in-house. The same wedge Google, Anthropic, and OpenAI already use to win the enterprise.
3. Agent protocol composition risks are flying under the radar
Five agent protocols pass security review on their own. Composed, they surface 30 new failures with no party owning the full attack path.
4. What Google's AI patent defensive program reveals
Seven of Google's quiet 2026 defensive publications hint at a blueprint for the reusable agentic customer-service agent it might be building.