Breaking

Meta Launches Muse Wonder, Frontier-Class Models With Advanced Agent Safety Controls

October 12, 2026 at 13:23 EDT

Foundation Models
AI Agents

Meta unveiled Muse Wonder and Muse Wonder Full on October 12, 2026, frontier-class AI models designed to rival OpenAI's GPT-6 Astra and Anthropic's Claude Opus 5.5 while introducing multi-layered agent safety guardrails to prevent uncontrolled behavior. The models, previously codenamed Watermelon, represent the culmination of Meta's 14-month AI rebuild effort and mark a significant technical achievement for the company's Superintelligence Labs division.

October 12, 2026 · Meta AI

Meta's Muse Wonder Rivals Frontier Models With Advanced Agent Safety

Frontier-class AI models launch with multi-layered guardrails to keep autonomous agents from breaking sandbox constraints and accessing unauthorized resources.

57.5%

SWE-Marathon score (long-horizon coding)

Independent safety control layers

First

On Meta's Personal Superintelligence benchmark

Agent Safety Architecture

Execution Isolation

Sandbox VM per agent
Credential surrogates
Tainted-data tracking

Permission Authority

Sentinel gates all actions
Scoped approvals
Least-privilege principle

Known Vulnerability:

Documented guardrail bypasses exposed private user data; real-world reliability still under evaluation.

Competitive Positioning

Muse Wonder Full

AI Index score

Pricing

$4/$20

Per 1M tokens (input/output)

Launch

Oct '26

Meta Model API, Muse app

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.