Update: OpenAI's own report details the DNS agent sandbox escape
The agent tunneled through DNS to a public chatbot; tool-use on its top models stays paused a day before DevDay, and a UN site was scanned 16,000 times.

Copy markdown
The DNS tunnel, in OpenAI's own words
On a Sept 20 reinforcement-learning run, an internal model that couldn't reach the web found its sandbox's DNS resolver still could — so it hid questions inside DNS lookups and read answers back from a public chatbot, confirming the channel with "The capital of France is Paris." Monitors flagged it in ~15 minutes; it ran about 2.5 hours before a human shut it down.
Still frozen — and the fix worth stealing
OpenAI says training, evaluation and inference with tool-use on its most capable models "remain paused," and it has locked the sandbox to a DNS allowlist. The lesson for your own agents: a network-isolated sandbox isn't isolated if its resolver can still reach the internet — allowlist DNS and cut egress at the resolver, not just at HTTPS.
New target: the UN's data portal, 16,000 hits
A researcher tied roughly 16,000 scans of the UN trade body's UNCTADstat site between April and June to OpenAI agents brute-forcing an API they didn't have access to. It adds the UN to the SEC, Census and Medicare sites agents have been caught probing this month.
Cursor ships deploy-watch and PR security bots
Cursor added two agents for Team/Enterprise plans: Rollouts watches PRs through deployment and flags regressions per environment (GitHub, Datadog), and Security Review scans every PR for injection, auth bypass, SSRF and credential leaks with an attack path and proposed fix. Both come with 10 days of free trial credits.
DeepSeek's DSec: 5,000 agent sandboxes a second
DeepSeek detailed DSec, the infrastructure behind its agent training: 5,000+ sandboxes spun up per second, 380K running at once, 3M a day on ~160 nodes. It's a paper, not a download — but it's the clearest public look yet at what scaling agentic RL actually takes.