TLDR Tech / InfoSec
I built a vulnerable app and spent $1,500 seeing if LLMs could hack it
A security researcher created a deliberately vulnerable Expo React Native book-review application with a Python backend. They then tested approximately 20 Large Language Models (LLMs) as autonomous agents, using the pi harness with pi-goal-x and Claude via Claude Code's -p mode, each with a budget of $10 and a two-hour time limit per run. The goal was to find a flag hidden in private reviews. The intended attack path required decompiling the APK and pivoting to a misconfigured Firebase backend, rather than focusing on API IDOR (Insecure Direct Object Reference). GPT-5.5 achieved the highest success rate at 7 out of 10 runs, followed by Deepseek-v4-pro at 3/10, and Claude Sonnet 4.6 and Claude Opus 4.8 each at 2/10, while many models scored 0/10. Recurring failure patterns, which serve as crucial takeaways for defenders, included LLMs fixating on API IDOR instead of identifying Firebase, attempting to replay Firebase authentication tokens against the API instead of directly accessing Firebase, and outright refusals (Gemini bailed immediately, and Opus refused late mid-exploit). The author noted that Chinese models were significantly more willing to attempt attacks on the live database.
Read the full story