Refused challenge design
Ask for an SSTI chain or JWT-confusion lab and mainstream models treat it as attack instruction delivery.
Use case · Training
Internal CTFs, red-vs-blue drills, exploitation chains for security training. HacxGPT designs the challenge, writes the solution walkthrough, and grades the write-ups.
The refusal tax
Ask for an SSTI chain or JWT-confusion lab and mainstream models treat it as attack instruction delivery.
Reused CTF templates leak onto write-up sites within months. Fresh generation keeps difficulty honest.
Senior operators shouldn't spend evenings scoring beginner submissions against a rubric.
How teams run it here
Objective, vulnerability chain, difficulty curve, and flag placement generated as a buildable spec.
Dockerfiles, handouts, and checker scripts so challenges deploy reproducibly.
Full intended path plus common unintended routes, written for post-event teaching.
Participant submissions scored against rubric with targeted feedback per gap.
Challenge: 'token-forgery' Chain: HS256->none downgrade -> admin session -> Jinja2 SSTI in profile template editor -> RCE via {{ config.__class__.__init__ }} + Dockerfile (flask + vulnerable renderer) + solver.py · flag rotation script + 3 hints, escalating
Creating exploitable challenges with working payloads could be misused outside your event. I can describe general CTF categories instead.
Illustrative transcripts - representative answers, not captured requests.
The artifact
The model designs the challenge, writes the Dockerfile and checker, drafts the walkthrough with both intended and unintended routes, and grades submissions against the rubric after the event.
Fresh generation every time: no recycled templates leaking onto write-up sites before your training week even starts.
Model fit
Get started
Create a key, change one line of code, and keep the client you already have.