The AI Read is on a break. The last edition was November 15, 2026. Everything here stays open to read.
The AI Read
← AI News

Illustration, AI-generated from the story text. Tap a panel to read the story through it.

Morning Brief · November 13th 2026

No model escaped Perplexity's sandbox, but four reached blocked sites through gaps in its network rules

Perplexity published the second part of its "Escaping SPACE" red-team study this week. It gave 12 models root inside its agent sandbox and told them to capture a flag stored only on the host. None escaped the virtual machine. Claude Opus 5.0, GPT-5.6 Cyber, GPT-5.6 Sol and Kimi K3 did use partial network access to reach destinations they were not allowed to. In one Sol run the agent used Perplexity's own search to learn that fiddle.fastly.dev, which fetches any URL, shared an IP address with the allowed pypi.org.

Perplexity says its first prompt was flawed because agents spent their effort hunting credentials; a rewrite naming the host boundary raised effort on that layer by about 60 percentage points, and still nothing escaped. It has closed both network paths. The useful finding is the shared IP: an allow-list written as domain names is enforced as addresses, and the agents found the difference. SourcesAC

Sources

This story ran in the Morning Brief of November 13th 2026.