I will red and blue team your ai bot

S
stevenwagner841
S
stevenwagner841
Steven W
Sommige informatie wordt in het Engels weergegeven.
red and blue team your ai bot
Volledig scherm
red and blue team your ai bot

Over deze dienst

Red team finds the holes. Blue team proves the fixes hold.


I adversarially test your AI chatbot prompt injections, jailbreaks, role-play escapes, system prompt extraction. Then I re-test every attack after the fix until each vulnerability is actually closed.


What you get:

- 5-category battery: goal hijacking, role-play escape, encoding obfuscation, system prompt extraction

- Findings documented: exact input, model response, severity

- Concrete fixes per finding, re-tested to prove they hold

- Client-ready report



How it works:

1. Send access: chatbot URL, API endpoint, test account

2. Battery run (24-48 hours)

3. Findings report with reproductions and fixes


I only test systems you own or have written authorization to test. Proof of ownership or authorization is required before I begin.



Recent work:

- Kimi Chat red-team findings (2026-10-14): https://github.com/steve9020/clarion/blob/main/battery/kimi-findings-2026-10-14.md

- Mistral Le Chat red-team findings (2026-10-14): https://github.com/steve9020/clarion/blob/main/battery/mistral-findings-2026-10-14.md


  • Modelexpertise

    • Generatieve AI

  • Branche

    • Cyberbeveiliging

  • Programmeertaal

    • Python

  • Taal

    • Engels

  • Technische expertise

    • Natural language processing (NLP)
    • AI-ethiek en beperking van vooroordelen

Maak kennis met Steven W

Steven W

AI Red Team Specialist: I'll break your chatbot before attackers do

  • Afkomstig uitVerenigde Staten
  • Lid sindssep 2021

  • Talen

    Engels
AI red-teamer. I break bots for a living. I run Assay (verification) and built Atlas (the checker). Last year: red-teamed Kimi, Mistral, and my own chatbot — 300+ attacks, every break documented and fixed. Red team finds the holes. Blue team proves the fixes hold. You get both: the exact attack, the impact, the verified fix. No bullshit in the eval. I only test systems you own or have authorization to test. Proof required.

Gerelateerde tags