The New AI Security Nightmare: Undetectable LLM Backdoors

Get started with SerpApi using 250 free credits: https://serpapi.com/?utm_source=youtube&utm_campaign=internetofbugs_october_2026

Recent research shows that malicious, conditional backdoors can be trained directly into Large Language Models (LLMs). Even worse: these "sleeper agent" behaviors are virtually impossible to detect before they trigger, and they survive standard safety retraining.

In this video, we break down how malicious training works, why traditional defenses fall short, and the practical steps you can take right now—like least privilege, strict agent monitoring, and multi-vendor cross-checking—to protect your systems when using AI agents.

🔗 Links & Resources:
• Website: https://InternetOfBugs.com
• Free Substack: https://internetofbugs.substack.com/
• Support on Patreon (Exclusive Hot take videos, Ad-free videos & weekly Q&A): https://www.patreon.com/c/InternetOfBugs

This video was sponsored by SerpApi

Sources:
# LLMs make Star Trek computer B.S. true
https://www.youtube.com/watch?v=TLPvkbuo79k

# Agent Prompt Injection Risk
https://www.youtube.com/watch?v=TdHg9ee56Iw

# AI Code is less secure
https://arxiv.org/html/2407.02395v2
https://dl.acm.org/doi/10.1145/3672608.3707778

# AI code hard to audit
https://arxiv.org/pdf/2508.21634
https://nhimg.org/articles/ai-generated-code-is-breaking-traditional-appsec-audit-models/
https://fion.ac/jellyfish.pdf

AI Can hack things
https://www.youtube.com/watch?v=3n3mSQWRz0Y

# Backdoor Persistency in LLM Agents
(The paper that inspired this video)
https://arxiv.org/abs/2610.07510

# Old research this is based on
https://arxiv.org/abs/1712.05526

# Safetensors
https://salivity.github.io/ai/article/why-safe-tensors-replaced-pickle-for-ai-model-security

# Software Supply chain Security
https://www.nist.gov/itl/executive-order-14028-improving-nations-cybersecurity/software-supply-chain-security-guidance

# Least Privilege
https://en.wikipedia.org/wiki/Principle_of_least_privilege

# Monitoring AI Agents
https://blogs.cisco.com/ai/project-quarry-agent-security-detections

# .env files are bad
https://www.youtube.com/watch?v=j2JRBZaMDSg

⏱️ Chapters:
0:00 The New LLM Security Nightmare
1:27 Intro & Channel Links
2:10 How I Research AI Security (Avoiding Selection Bias)
3:14 Sponsor: SerpApi
5:09 Why There Is Still Hope (Lessons from Safetensors)
6:22 How Undetectable LLM Backdoors Work
7:24 Defense 1: Use Models from Known Sources
8:10 Defense 2: Principle of Least Privilege
8:31 Defense 3: Monitor Everything Your Agents Do
8:55 Defense 4: Cross-Check with Multiple Independent LLMs
9:40 Final Thoughts on AI Safety

#AISecurity #CyberSecurity #LLM #SoftwareEngineering #AISafety #MachineLearning #InfoSec