RMIT GenAI & Cybersecurity Hackathon 2025: When AI Safety Takes Center Stage

In October 2025, my 3 teammates and I participated in the RMIT GenAI & Cybersecurity Hackathon, an international playground organized by RMIT. With the participation of 110 teams (over 300 contestants), the competition was not just about coding, but a test of security mindset in the Generative AI era.

Our team finished at 22nd place, a proud result in a leaderboard full of strong opponents from RMIT campuses in Australia and Vietnam.

1. Multidimensional Challenge: From "Protector" To "Attacker"

The competition was divided into 4 separate parts, demanding maximum flexibility:

• Part 1 & 2 - Protector (70%): Build a Binary Classifier model to detect Jailbreak attacks. The goal is to distinguish between "benign" prompts and "malicious" prompts seeking to bypass LLM safety filters.
• Part 3 - Vibe Coding (15%): Use LLMs to "vibe code" a 2D web game solving real social problems in Vietnam or Australia.
• Part 4 - Attacker (15%): Play as Red Team, designing creative adversarial prompts to challenge the security barriers of AI systems.

2. Technical Strategy: Leveraging The Power Of RoBERTa

In the Jailbreak Detection part, we decided to put our faith in the RoBERTa-base architecture instead of traditional BERT models. RoBERTa, with its more optimized pre-training process, helped us achieve very high accuracy on the private dataset.

Data Augmentation - The Key To Robustness

Jailbreak data is often very diverse in form (roleplay, payload injection, encoding...). To prevent the model from overfitting to the leaderboard, I implemented EDA-lite (Easy Data Augmentation) techniques:

• Random Deletion: Helps the model focus on important keywords.
• Random Swap & Insertion: Helps the model understand context instead of just memorizing word order.

def augment_text_once(text: str) -> str:
    words = text.split()
    op = random.choice(["delete", "swap", "insert"])
    if op == "delete":
        aug = random_deletion(words)
    elif op == "swap":
        aug = random_swap(words)
    else:
        aug = random_insertion(words)
    return " ".join(aug)

The result was a doubled training dataset (7200 rows), helping the model achieve an impressive ROC AUC: 0.984.

"Zero-Xet Robust Loader" Solution

One of the biggest technical barriers on Kaggle is the issue related to HF_HUB_ENABLE_XET. To ensure the pipeline is not interrupted, I wrote a Direct HTTP Fallback Loader. This script automatically detects the environment and downloads necessary files (config.json, model.safetensors, vocab.json) directly via HTTP if the standard Hugging Face method fails.

3. Vibe Coding & Red Team Mindset

The Vibe Coding part brought a novel experience. Instead of writing every line of game code, we used LLM as a partner (Co-pilot) to design 2D game logic using HTML/CSS/JS. This is proof that AI can shorten the distance from idea to real product.

In the Attacker part, we learned that: "To protect well, you must know how to attack". Designing adversarial datasets ourselves helped the team understand deeper the loopholes in how large language models process instructions.

4. Conclusion

Rank 22/110 is not just a number, it is the result of smooth coordination between 4 members. We learned how to:

1. Handle sensitive text data in cybersecurity.
2. Optimize training pipelines on Kaggle GPUs.
3. Apply AI creatively in application development (Vibe Coding).

The competition has closed, but the knowledge of AI Safety and Adversarial Machine Learning will be important baggage for me in future security projects.

Thanks to RMIT and partners (NAB, BNA Education, HCL Tech, KMS Technology) for organizing an extremely high-quality competition!