Zero Theory Labs

Cybersecurity for the age of AI attackers.

AI agents now run intrusions and code changes in seconds, with nobody steering. Zero Theory Labs builds security tools for that speed — testing every claim against recorded evidence rather than shallow checks.

Every check ends one of three ways.

A cube of small blocks: most are solid, a few are coral, some are dashed outlines

Security checks pass. Attacks finish anyway.

Two things are true at once: the checks teams rely on test the wrong thing, and autonomous agents have stopped waiting for people.

Shallow checks miss real blast radius

Liveness probes and unverified static alerts ask whether a process responded or a keyword matched — not whether untrusted input actually reaches a dangerous sink or critical state stayed intact.

Attacks now run themselves

AI agents find a weakness, exploit it and move on with nobody steering. The techniques are old. The speed is new.

2–25 s How long one burst of agent activity lasts
27 in 23 s Events in one burst, testing a stolen cloud key
240 Log events from one attempt to run code

Figures reported by Google and Mandiant, 30 October 2026.

Marlowe · v0.4 early access

A terminal with a security workbench beside it.

Marlowe pairs real shells with a taint-verifying security agent, nine local security CLIs, Shannon-entropy Codebase DNA and attack-reachability maps. Every overwrite, deletion and command waits for your approval before it runs.

Watch the live build & explore Marlowe
  1. Scoped Lock network recon to authorized hosts and auto-detect Semgrep, Trivy, Gitleaks, OSV-Scanner, Nmap, Nuclei, TShark, Radare2 and otool.
  2. Mapped Inspect Codebase DNA byte entropy, 9-CLI multi-engine convergence and reachable vs. severed false-positive paths.
  3. Approved Overwrites, deletions and commands pause at a host-enforced line-diff gate. One approval covers one action.
  4. Verified Connect local Ollama/vLLM or Claude, OpenAI, Gemini and Grok. Export verified findings as Markdown, SARIF v2.1.0 or JSON.

Nothing is claimed here until a recorded run shows it.

Zero Theory Labs is a one-person lab that builds tools to test whether systems do what they claim. Additional containment and recovery tracks remain in sealed range development while Marlowe. (v0.4) opens for invite-only early access on macOS (with experimental Windows and Linux builds).

Jessie Hermosillo
Principal SRE, founder

  • Three answers, never two Every check reports holds, fails or can't tell. Unknown is never rounded to a pass.
  • Proof before claims A claim is made only after its scenario passes on a recorded run.
  • Limits stated up front Targets are labelled as targets and examples as examples. What a tool does not do is written next to what it does.

Ask for early access to Marlowe.

Enter your email and GitHub handle to receive a Read-only invite to zerotheory-labs/marlowe (v0.4 · PolyForm Noncommercial 1.0.0).

Prefer email? Write to access@zerotheorylabs.com · Watch the live build →