Meta launches Muse personal agent with OS-level defenses against prompt injection
DeepLearningAI · x · 2026-09-22
Meta unveiled Muse, a personal AI agent powered by its Muse Spark 1.3 model, accessible via app or WhatsApp. It reads and sends email, browses the web, fills forms, makes purchases, and keeps running in the background. Pricing: free (100M tokens/week), $20 (500M), and $100 (3B); US-only for users 18+.
Security is the headline feature: Muse assumes the model will be prompt-injected, so defenses live at the OS level — the model never handles real credentials, tools run in isolated Linux containers, and an independent gatekeeper verifies outbound calls. Model architecture, training data, and safety evals remain undisclosed.
More from coding & agent
- Agentic LLMs Can Write Rust Code 2x-20x Faster Than State-of-the-Art Libraries — minimaxir · 2026-09-22
- Aethos X hits 94% on EnterpriseRAG-Bench at 42k files, nearing effective ceiling — rohanpaul_ai · 2026-09-22
- Grok 4.7 generates a playable Plants vs Zombies from a single prompt — vista8 · 2026-09-22
- LlamaIndex adds per-field confidence scores to Extract for human-review triage — llama_index · 2026-09-22
- LangChain's Jev judge matches Claude Sonnet 4.6 evals at $0.34 vs $28.17 per run — LangChain · 2026-09-22
- LangSmith ships Jev-as-a-judge to score every production agent trace at low cost — LangChain · 2026-09-22