Netrunners
Xploit is for game developers and QA teams testing economies with virtual currency, rewards, and item upgrades. Manual testing often misses exploits that require several actions in sequence, such as resetting a profile to claim rewards repeatedly or buying, upgrading, and reselling items for unintended profit. These loopholes can undermine progression and player fairness. Teams need a faster way to discover these exploits, reproduce them reliably, and verify that fixes work before release.
Preliminary score
- Value for the user18.5 / 25
A specific problem and a clear outcome from solving it.
- Prototype and use of AI22.0 / 30
A working core scenario, and what AI actually contributes to the result.
- Quality testing16.5 / 20
Tests, examples of failures, and comparison with the current approach.
- Feasibility10.5 / 15
Data requirements, running costs and a clear next step.
- Originality7.0 / 10
A meaningful difference from common solutions.
What the judges wrote
- GPT-6 AstraPreliminary78/100
- Strengths
- Your design gives AI a concrete exploration role while reserving exploit confirmation for an independent deterministic verifier. Your reported reward-reset and resale findings include fresh-session replay and secure-control checks, giving QA teams reproducible evidence rather than model assertions. Your evaluation earns credibility by retaining failed runs and acknowledging that the random baseline also finds exploits.
- To improve
- Your next benchmark should use held-out economies, more repeated trials, and matched budgets against random and conventional stateful testing to establish where AI adds value. Validate one developer-provided game adapter with a QA team and measure integration effort and testing time saved. Add per-run token costs, latency summaries, and a tested quota-recovery strategy rather than relying on free-tier availability.
- Claude Opus 5.5Preliminary71/100
- Strengths
- Your problem framing is sharp: multi-step economy exploits such as reward reset and buy-upgrade-resell loops are a real blind spot for endpoint-level QA, and you show both with concrete coin balances. Your architecture separates concerns well. The LLM only chooses actions, while a deterministic verifier, a fresh-session replay and a secure-control comparison confirm findings, so results are evidence rather than model opinion. Your testing is notably honest. You have 51 automated tests, 30 logged trials against a seeded state-aware random baseline, retained failed and quota-interrupted runs, and you state plainly that AI superiority is not established.
- To improve
- Your own baseline found the exploits as often as Gemini did (2/2 vs 1/2 on reward reset), so you have not yet shown why AI is essential. Build a scenario where random search provably fails, such as a long or conditional action chain, and report discovery rate and steps-to-find over many more trials. Measure actual cost per run (tokens, calls, latency) instead of 'cost not measured', and add a comparison with how a QA engineer finds the same bug today. Your scenarios are deliberately injected into a toy sandbox, so testing one unseen, held-out economy, or a developer-provided adapter, would greatly strengthen the value claim.
The team
Huseyn Pashayev
data
Farida Orucova
developer
Ismayil YusifliLead
data