About this site

Adversarial Logic is a technical blog on adversarial machine learning, LLM security, and the security of AI agent systems. Posts here are research-grounded, claims are cited, code examples are tested, and tradeoffs are stated honestly. The goal isn't to sell you a product or hype the latest threat. It's to explain what the research actually shows about how ML systems break and what works (and what doesn't) to defend them.

What to expect?

Every post on this site follows a few rules:

  • Claims cite sources. Papers, CVEs, security advisories, or documented incidents. If a number appears in a post ("70.97% attack success rate"), the citation tells you exactly where it came from.
  • Code examples are tested. If something only works on CIFAR-10 and falls apart at higher resolution, the post says so. Caveats and failure modes are stated, not buried.
  • Tradeoffs are stated. No defense gets described as a silver bullet. Posts try to make the limitations as clear as the wins.

The premise is that engineers and security practitioners deserve writing that takes the research seriously without burying it in jargon or hype.

Who writes this stuff?

Written by Josh, a computer scientist with a Master's in machine learning security and active experience in adversarial ML. Independent — not sponsored, not affiliated with any AI security vendor.

The pseudonymous byline is intentional. The work stands on the citations, the code, and the technical reasoning. All of it is on the page.

Topics you can expect

  • Adversarial examples in computer vision (one-pixel attacks, sparse perturbations, vision transformer vulnerabilities)
  • LLM jailbreaks and prompt injection (evolutionary methods, gradient attacks, defense evaluation)
  • RAG security and document-level injection
  • AI agent threat modeling and architectural defenses
  • Supply chain attacks on ML systems
  • The robustness and alignment literature

Start Here

If you're new to the blog, these posts cover the strongest material:

Corrections and contact

If you find an error in a citation, a broken code example, or a claim that doesn't hold up, email josh@adversariallogic.com. Corrections will get posted in a timely manner.