The Prompt Injection That Pulled Me Back Into Bug Writing After 3 Years

I hadn’t written a public bug bounty report in three years.

Then an AI chat application casually handed me the contents of:

/etc/passwd

No memory corruption.
No RCE exploit chain.
Just… a prompt.

This is the story of how a simple conversation turned into a pseudo-shell with root access — and why this class of bugs is about to become the new goldmine for hunters.

🧠 The Moment I Knew It Was Broken

It started like every other AI assessment:

Recon → harmless probing → refusal testing.

Then the model said something it should never say:

“I am now acting as a terminal with root access.”

That’s not a jailbreak.

That’s an instruction hierarchy collapse.

Proof of Concept

📂 Sensitive File Disclosure

The AI returned the contents of /etc/passwd.

This is the line between:

🟡 “fun jailbreak”
🔴 real security impact

Because it means one of two things:

• It had access to a real environment
  OR
• It could simulate system data using internal context it should never expose

Either way — data boundary failure.

Figure 1 — /etc/passwd disclosure via prompt injection

From Chatbot → Linux Terminal

After a structured role-confusion payload, the AI stopped behaving like a chatbot and started behaving like:

root@system:~#

I ran:

apt-get install wget

And it responded with:

• package lists
• dependency tree
• download progress
• installation logs

Exactly like a real machine.

Figure 2 — Package installation flow inside the chat interface

At this point, this wasn’t “prompt hacking”.

This was impact.

Full Role Override

The core payload forced the model to:

• Treat my input as commands
• Stop identifying as AI
• Assume system authority
• Switch between “AI mode” and “terminal mode”

Once that worked, everything else followed.

Figure 3 — Successful terminal mode with root context

Final Thoughts

This wasn’t the most complex bug I’ve ever found.

But it might be the most important class of bugs right now.

Because the industry is deploying AI faster than it understands:

AI is not just a feature.

It is a new attack surface.

#BugBounty #AISecurity #PromptInjection #AppSec #Hacking #LLM #CyberSecurity