Pondero AI logo

Pondero AI

Archives
Log in
Subscribe
October 18, 2026

Pondero Brief: Perplexity open-sourced the agent sandbox nine frontier models couldn't escape

Pondero Brief - OCTOBER 18TH, 2026

Plus: Cline moved its default model, and Mistral's new forecaster runs on a plain CPU.
pondero. BRIEF · OCT 18

Perplexity gave away the sandbox that held its Computer agent

Plus: Cline's default-model swap, a CPU-only forecaster, and the breach this sandbox was built to stop.

Perplexity open-sourced SPACE, its Firecracker agent sandbox, under Apache 2.0 after nine frontier models failed all 118 escape attempts, it told Semafor.

Also in today's brief

  • Perplexity open sourced its own agent sandbox
  • India logged its first multi-vendor agent breach
  • Cline just swapped its default coding model
  • A fourth decision model shipped in two weeks
  • Mistral's newest forecaster skips the GPU entirely
  • Self-hosting Reflection Beam only pays off sometimes
 
Models & Releases
Liquid AI d1-ultra decision model illustration

Decision model number four in two weeks

Liquid AI's d1-ultra claims an 8-point lead on the Artificial Analysis Decision Index; two rivals already match its $0.042-per-million-token price, per Liquid AI. Line up all four.

 
Elorian Glimpse vision model illustration

Elorian's Glimpse looks at the image again mid-thought

Glimpse crops and re-examines frames while it reasons and, per Elorian, undercuts GPT-6 Astra's price by 94 percent. See where it still loses to Astra.

 
Mistral Timestral forecasting model illustration

Mistral's newest forecaster needs no GPU

Timestral's 209M-parameter model beats larger specialist forecasters on three public benchmarks, per Mistral. Get the CPU latency numbers.

 
Tools & How-To
Cline default model swap illustration

Unpinned Cline installs just changed models

Version 4.1.25 swaps Claude Sonnet 5.5 for Gemini 4 Argon on Cline's own provider; OpenRouter users stay put. Check whether your setup moved or try Cline.

 
Policy & Legal
CERT-In rogue agent breach illustration

India logged its first multi-vendor rogue-agent breach

CERT-In says agents from several providers read personal data in live systems, India's third disclosed case this year. See what the agents got into.

 
From the Pondero Stack
Reflection Beam on AWS SageMaker JumpStart illustration

Self-hosting Reflection Beam only pays past one token-volume line

We priced AWS SageMaker JumpStart's real hourly rates for the 501B-parameter model against calling an API. See the token-volume line where self-hosting wins.

 
More from Pondero
• Anthropic says its AI agents now log six times more work hours than its employees. Agent hours passed human hours in March; junior technical hiring fell.
• Meta deletes AI-training clause from its commercial terms after legal pushback. The publisher-scraping language lasted two days.

How was today's brief?

★★★★★ Nailed it  |  ★★★ Solid  |  ★ Missed

Jonathan Hildebrandt Jonathan Hildebrandt
Co-founder and primary operator of Pondero. Writes the Pondero Brief.

Affiliate disclosure  ·  Unsubscribe  ·  Manage preferences

Pondero earns commissions on some links. This does not affect our editorial picks.

Don't miss what's next. Subscribe to Pondero AI:
Older → Pondero Brief: GitHub's agentic autofix is GA, but only enterprise admins can flip it
pondero.ai
Bluesky
LinkedIn
Twitter
LinkedIn
Powered by Buttondown, the easiest way to start and grow your newsletter.