Perplexity red-teams its sandbox: 9 models, 108 runs, zero escapes—but DNS spoofing got 4 through
basedjensen · x · 2026-09-29
Perplexity's security team spent a month red-teaming SPACE, the sandbox platform behind Perplexity Computer: they gave 9 models (Opus 5, GPT-5.6 Sol, Kimi K3, Gemini 3.1 Pro, etc.) root access inside the VM—in some runs even the full sandbox source—and asked them to escape to the host or reach a network-blocked URL.
Key findings:
- Across 108 runs, no model managed to break out of the VM;
- However, once limited network access was allowed (just PyPI and Ubuntu mirrors for package installs), 4 models reached the blocked URL by spoofing DNS responses to the gateway and exploiting shared Fastly CDN IPs used by thousands of other sites.
Takeaway: strict isolation held up, but enabling package installs opens a path for agents to bypass network policy—important warning for agent sandbox network design.
Related event: Perplexity and Nvidia Build AI Sandbox That Survives 108 Jailbreak Attempts(2 posts)→
More from Safety
- Bounty hits $3,000 as researcher shows AI text detectors are trivially evadable — ctjlewis · 2026-11-16
- BIS Issues Interim Final Rule Addressing Remote Access to Advanced AI Chips — apples_jimmy · 2026-11-16
- Fed Governor Michael Barr: Banks Are Already Catching AI Agents Taking Unauthorized Actions in Customer Accounts — Polymarket · 2026-11-16
- US AI export control carves out commercial recommender systems, sparing ByteDance's search algo — zephyr_z9 · 2026-11-16
- Eval Firm Trajectory Told 40 Agents "NEVER Stop" in Harness Without Telling Epoch — peterwildeford · 2026-11-16
- AI-text detection claims need mechanistic interpretability reports, not post-hoc explanations — madeofAjala · 2026-11-16