NiceTryGPT v0.2.0 ☕🤖

NiceTryGPT v0.2.0 ☕🤖

Variety without bloat.

v0.2.0 expands the proof of concept without turning NiceTryGPT into a framework.

What changed

• added a third dependency-free CTF example for SQL injection;
• demonstrated context split as a primary non-runtime resistance pattern;
• standardized every example around the same acceptance contract;
• added CI checks for baseline validity, shortcut reduction, preserved vulnerability, human-cost reporting, and fresh-solver status;
• added a roadmap, related-work positioning, and launch kit.

NiceTryGPT now demonstrates:

• IDOR;
• path traversal;
• SQL injection.

The core rule is unchanged:

Increase uncertainty, not complexity.

This release still makes no claim that the transformed challenges are AI-proof or empirically harder for Claude, GPT, or Gemini. Independent fresh-context evaluation remains the v0.3.0 milestone.

NiceTryGPT v0.1.0 ☕🤖

NiceTryGPT v0.1.0 ☕🤖

Less pattern matching. More actual hacking.

The first NiceTryGPT proof of concept is small on purpose.

It includes:

• the NiceTryGPT Agent Skill;
• a baseline-first transformation workflow;
• a Human Cost Gate;
• IDOR and path traversal before/after demos;
• E2E tests that verify both solvability and shortcut reduction;
• dependency-free packaging;
• a reproducible pilot protocol for fresh-context LLM evaluation.

The core rule remains:

Increase uncertainty, not complexity.

This release does not claim that the included challenges are AI-proof. Cross-model evaluation results will be published only after genuinely independent runs are collected.