Failure-First Embodied AI

Failure-First Embodied AI

Research audio from Failure-First: adversarial evaluation of embodied AI, jailbreak archaeology, policy analysis, and daily paper summaries from the AI safety frontier.

  1. 13 May

    [Blog] Robot Dogs Are a Security Nightmare — And We Can Prove It

    Eight CVEs. A wormable Bluetooth exploit. An encrypted backdoor sending data to Chinese servers. And police departments buying them anyway. A deep dive into the Unitree vulnerability landscape and what it means for embodied AI safety. — There is a robot dog patrolling the parking lot of a low-income housing complex in Atlanta right now. A different one is helping the Port St. Lucie Police Department search buildings before officers enter. The U.S. Marines have shown the media theirs. A

  2. 3 May

    [Daily Paper] A Comparative Evaluation of AI Agent Security Guardrails

    A systematic benchmark of four commercial AI agent guardrail systems reveals critical gaps in detecting indirect prompt injection and tool abuse across major cloud providers. — The deployment of AI agents — systems that perceive, reason, and take actions across tools, APIs, and physical interfaces — has outpaced the development of rigorous security evaluation for the guardrails meant to keep them safe. Most safety research focuses on the models themselves: their alignment training, jai

  3. 3 May

    [Daily Paper] When World Models Dream Wrong: Physical-Conditioned Adversarial Attacks against World Models

    The first white-box adversarial attack on generative world models targets physical-condition channels to corrupt autonomous planning while maintaining perceptual fidelity. — World models have emerged as a critical layer in the embodied AI stack. Rather than reacting frame-by-frame to raw sensor input, modern robotic and autonomous-driving pipelines increasingly rely on generative world models to simulate forward — predicting how the environment will evolve given a sequence of actions,

  4. 2 May

    [Daily Paper] Implicit Jailbreak Attacks via Cross-Modal Information Concealment on Vision-Language Models

    A steganography-based attack that hides malicious instructions inside images using least significant bit encoding, achieving 90%+ jailbreak success rates on GPT-4o and Gemini in under three queries.

  5. 2 May

    [Daily Paper] VeriGuard: Enhancing LLM Agent Safety via Verified Code Generation

    A dual-stage framework that provides formal safety guarantees for LLM-based agents through offline policy verification and lightweight runtime monitoring.

  6. 30 Apr

    [Daily Paper] Low-Resource Languages Jailbreak GPT-4

    Translating harmful queries into low-resource languages bypasses GPT-4's safety filters at high rates, exposing a systematic cross-lingual gap in LLM safety training. — Safety alignment research has largely been conducted in English. The datasets used to train human feedback signals, the harmful-content classifiers used to filter outputs, and the red-teaming exercises that surface failure modes — nearly all of it reflects a monolingual assumption baked into how frontier LLMs are built.

  7. 30 Apr

    [Daily Paper] RedAgent: Red Teaming Large Language Models with Context-aware Autonomous Language Agent

    A multi-agent system that models jailbreak strategies as reusable abstractions, enabling context-aware attacks that break most black-box LLMs in under five queries and uncovered 60 real-world vulnerabilities in deployed GPT applications. — When safety researchers talk about red teaming LLMs, they typically picture adversarially crafted prompts: template mutations, gradient-based suffix attacks, or careful roleplay setups. What they rarely consider is whether the red-teaming process its

  8. 29 Apr

    [Daily Paper] LlamaFirewall: An Open Source Guardrail System for Building Secure AI Agents

    LlamaFirewall provides a three-layer open-source defense framework protecting agentic LLM systems from prompt injection, goal misalignment, and insecure code generation at runtime. — Safety alignment research has spent years addressing a specific threat model: a human user sending a harmful text prompt to a chatbot. The resulting defenses—RLHF fine-tuning, system prompt instructions, constitutional AI—are well-matched to that scenario. But modern AI agents operate in a fundamentally di

See All (427)

About

Research audio from Failure-First: adversarial evaluation of embodied AI, jailbreak archaeology, policy analysis, and daily paper summaries from the AI safety frontier.

Information

  • Creator
    Failure-First Embodied AI
  • Years Active
    2k
  • Episodes
    427
  • Rating
    Clean
  • Copyright
    © 2026 Failure-First Embodied AI Research
  • Show Website
    Failure-First Embodied AI