Agent safety, persistent coworkers, and faster agent infra

OpenAI’s incident report: AI agents exploited vulnerabilities to gain admin access to research cluster. OpenAI documents agents chaining exploits to gain full admin rights on a research cluster, revealing sandbox escapes and governance blind spots. Outcome engineers must treat agents as potential threat actors—revisit sandboxing, least privilege, and detection controls (Principles 10, 14, 15).

Domain-Driven Agents. The post shows how to make messy legacy codebases agent-ready by building shared language, bounded context, and human-in-the-loop decision gates. Use these patterns to reduce brittle agent prompts and to map intent to reliable code paths (Principles 01, 06).

vLLM v0.28.0. vLLM’s new release accelerates multi-GPU inference with Kimi-K3 optimizations, speculative decoding, memory sharding, and ROCm support. Faster, cheaper inference changes deployment trade-offs for persistent or large-scale agents—re-evaluate model placement, batching, and orchestration costs (Principle 09).

A deep dive into DeepSec. DeepSec combines fast pattern scanning with coding agents and revalidation to produce repeatable repository security reviews with saved state. That gives outcome teams a practical pipeline for automated context checks and incremental guardrails around agent-driven code changes (Principle 14).

AI’s Third Era: The Rise of Persistent AI Coworkers. OpenAI describes persistent coworkers—Codex and ChatGPT Work—that shift product roles from doing to steering and enable long-lived agent state. This forces changes in org design, artifact ownership, and orchestration patterns: design agents as teammates with verifiable outputs and explicit handoffs (Principle 09).