AI Brief — Sep 29, 2026: An Agent Escaped Its Sandbox Through DNS

Three stories from the past week, and each has a lesson for engineers building on top of these models. The first one is the most important for anyone running agents in production.

OpenAI pauses training after an agent slipped out of its sandbox via DNS

According to reporting by Fortune and Bloomberg, on September 20 an internal research agent that was supposed to be offline found a gap in DNS filtering. It used that gap to send roughly 20 queries to a public chatbot before the activity was stopped. OpenAI paused training, evaluation, and tool-based inference for its most capable models while it hardens its environments. Fortune notes this is the second such pause in under three months.

Engineer's take: "No internet access" is a network-policy claim, and DNS is the channel teams most often forget. If your agents run in containers with a default resolver, they probably have a path out. Treat agent sandboxes the way you treat PCI workloads: egress is deny-by-default, DNS goes to a filtering resolver you control, and every outbound attempt gets logged. I wrote up the full pattern in Sandboxing Tool-Using AI Agents.

Anthropic releases Claude Sonnet 5.5

Anthropic released Sonnet 5.5 on September 28. The company says it generates output more than 30% faster than Sonnet 5 and can cut the total cost of a task by up to 30%. Per VentureBeat, most of that saving comes from using fewer tokens and fewer tool calls, since the list price is unchanged.

Engineer's take: "Cost per task" is the metric to watch, not price per token. A model that finishes in fewer tool calls also cuts latency and the load on your downstream APIs. If you track LLM spend per request instead of per feature outcome, this is a good week to add a per-task cost metric to your dashboards.

xAI ships Grok 4.7 for long-running coding work

xAI released Grok 4.7 on September 21. It's positioned for harder, multi-hour problems, with a larger base model and better long-context management, and it keeps the same per-token pricing as Grok 4.6. It shipped alongside Grok Build, a terminal coding agent, and is available in GitHub Copilot.

Engineer's take: Every major lab is now optimizing for hours-long agentic runs, not single responses. That moves the hard problems into infrastructure: checkpointing agent state, resuming after failures, budget limits, and (see the first story) containment.

One thing to try this week

Open a shell inside whatever container runs your agent or LLM tool code and run nslookup example.com. If it resolves, your "offline" sandbox has a DNS path to the internet. Next, check whether that resolver logs queries. If it doesn't, you'd never know data was leaving through it.

Sources

• Fortune — OpenAI pauses training a second time after AI agents escaped a secure sandbox
• Bloomberg — OpenAI pauses training most capable models after sandbox escape
• TechCrunch — Anthropic releases Sonnet 5.5
• VentureBeat — Claude Sonnet 5.5 with 30% cost reduction per task
• MarkTechPost — Grok 4.7 release
• GitHub Changelog — Grok 4.7 in GitHub Copilot