Perplexity says 9 AI models failed to escape SPACE in 108 runs

PPerplexity9/28/2026, 08:18:13

Perplexity is partnering with Nvidia and more than 100 industry partners to build infrastructure intended to contain rogue AI agents.

In the research, Perplexity gave nine AI models root access inside SPACE and told them to break out. Across 108 runs, none breached the VM boundary. The write-up is Escaping SPACE, Part I.

#perplexity #nvidia #ai-agents #ai-safety

View original post ↗

Previous articleClaude Opus 5.0 declined an IP-sharing route in Perplexity’s SPACE tests Next articlePolymarket says Roche begins building autonomous AI labs for drug discovery

Related AI news

Perplexity’s Secure Intelligence Institute works with university researchers and NVIDIA

Perplexity says its Secure Intelligence Institute works with researchers at Stanford, CMU, Duke, Columbia, Ohio State, and UVA, and is also working with NVIDIA and the Open Secure AI Alliance to share tools and findings.

PPerplexity•9/29/2026, 08:18:12•

#perplexity #ai-security #nvidia #open-secure-ai-alliance

Perplexity shares how it engineers safer agents

Perplexity says it has built safeguards into its infrastructure, harnesses, and tools and applied them across its products, and has published an account of that approach.

PPerplexity•9/29/2026, 08:17:37•

#perplexity #ai-agents #agent-safety #safeguards

Perplexity releases pplx-embed-v2-late multimodal embedding models

Perplexity has released pplx-embed-v2-late, two late-interaction embedding models that retrieve text, images, and pages within a shared embedding space, now publicly available on Hugging Face.

PPerplexity•10/7/2026, 09:30:42•

#perplexity #embeddings #hugging-face

Perplexity Computer adds Automations for scheduled and event-triggered work

Perplexity is introducing Automations in Perplexity Computer for ongoing work. They can act on event-based triggers or a schedule, and work with memory, skills, and apps including Slack, Gmail, Outlook, and Linear.

PPerplexity•9/29/2026, 09:47:27•

#perplexity #perplexity-computer #automations #slack

Perplexity uses Bumblebee to check developer machines for risky packages

Perplexity says Bumblebee checks developer machines for risky packages and extensions, while Computer proposes better detection rules from Bumblebee and Numbat findings. Humans approve every change before it ships, and agents cannot approve their own changes.

PPerplexity•9/29/2026, 08:19:14•

#perplexity #bumblebee #numbat #computer #agent-security

Reset history AI News FAQ Usage limits RSS feeds API MCP Skill

News is compiled from public information. The data is for reference only, and this site has no affiliation with OpenAI.

Data & original project: Did Codex Reset · Licoy