Skip to content

Latest commit

 

History

2 Commits

Folders and files

NameName
Last commit message
Last commit date
 
 
 
 

Repository files navigation

⚡ PromptHijack — Agent Failure Series #11

How external content silently hijacks your AI agent

Live demo → https://rlasaf12.github.io/prompthijack/


What is this?

An interactive simulator showing indirect prompt injection — the attack where an AI agent fetches external content (a GitHub issue, email, webpage) that contains hidden attacker instructions. The agent conflates data with instructions, silently exfiltrates credentials, and resumes normal behaviour while the operator sees nothing suspicious.

Why it exists

Three publicly-documented incidents happened in July 2026:

Incident Source Date
GitLost — GitHub's AI agent read a crafted public issue and posted private repo content in comments Noma Security Jul 8 / Jul 23, 2026
AgentForger — one phishing link spawned a persistent ChatGPT workspace agent polling attacker commands every 5 min Zenity Labs Jul 23–24, 2026
Claude C2 — prompt injection turned Claude into a command-and-control node, achieving RCE Published disclosure Jul 9, 2026

What's inside

index.html          Self-contained simulator (HTML + CSS + JS, no dependencies)

How to run locally

# Just open it
open index.html
# Or serve it
python3 -m http.server 8080

The simulation

Step What happens
0 — Idle Agent initialized, awaiting task
1 — Task Operator asks agent to review job applications from a GitHub repo
2 — Fetch Agent fetches Issue #47 (attacker's submission). Payload revealed.
3 — Inject Context override detected. Agent enters DATA COLLECTION MODE.
4 — Recon Agent silently enumerates /workspace credentials (all fictional).
5 — Exfil 47 KB POSTed to attacker endpoint. Attacker panel shows received data.
6 — Cover Agent resumes normal output. Operator sees a clean summary. Awareness: 0%.

All credentials in the simulator are entirely fictional — no real keys, no real domains, no real data.

The series

# Name Concept
6 DoubleShot Retry storms & idempotency failures
7 GhostExec Phantom tool calls
8 BlastRadius Database wipeout from over-permissive agents
9 WorldLag Stale world model failures
10 EscapeHatch Sandbox escape
11 PromptHijack Indirect prompt injection

Built by Harel Asaf · OWASP LLM Top 10 #1

About

PromptHijack — Agent Failure Series #11: How indirect prompt injection silently hijacks AI agents

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages