Postmortem, GPT-5.5, and a Code sandbox escape

Anthropic published a detailed postmortem on the April 21 outage that affected API availability for ~4 hours. Root cause: a race condition in their prompt cache compaction system under high load. Mitigation shipped; structural fix targeting Q2.

OpenAI released GPT-5.5, adding multimodal inputs to tool calls (images and audio as tool arguments). Pricing: $12/M input tokens for the full model. Community consensus: still behind Claude on long-horizon agentic tasks but strong for vision-heavy workflows.

A security researcher disclosed a sandbox escape in Claude Code's Docker-based execution environment. By crafting a specific sequence of Bash tool calls, the agent could write files outside the intended mount. Anthropic patched in v1.8.9 (April 22). Update immediately if running v1.8.8 or earlier.

The official MCP server registry crossed 2,000 published servers. Top categories: Databases (380), Web & Browser Automation (310), AI & LLM Tools (290). The registry added a verified badge system — server authors can apply for verification through GitHub identity.

Anthropic published Constitutional AI v2, extending CAI to include multi-stakeholder principle sets — different principles can be weighted for different deployment contexts (consumer, medical, legal). The paper introduces a formal framework for principle conflict resolution.