To be an effective AI Safety Researcher, you must look beyond simple bugs. The landscape of AI vulnerabilities is complex, spanning technical glitches, systemic failures, and ethical breaches. At Genbounty, we conceptualize this as the AI Safety Stack - three distinct layers of defence that every researcher must understand.

1. The Tactical Layer: OWASP Top 10 for LLMs

The Hacker’s View (Security & Exploitation)

The OWASP Top 10 for LLMs is your battleground. It defines specific, technical failure modes that allow an attacker to compromise a system. This is the 'Target List' for offensive testing.

Focus: 'Can I break it? Can I subvert the controls?'
Key Risks (2025 Edition):
- LLM01 Prompt Injection: The art of overriding the System Prompt to force unintended behaviors (Jailbreaking).
- LLM02 Insecure Output Handling: When the AI generates code or content that is executed by a browser or backend without validation (leading to XSS or RCE).
- LLM06 Excessive Agency: Tricking the AI into taking unauthorized actions, like deleting files or sending emails.
Genbounty Map: These findings fall under Product-Level Vulnerabilities. They are usually high-severity technical findings.

2. The Process Layer: NIST AI RMF

The Engineer’s View (Reliability & Lifecycle)

The NIST AI Risk Management Framework isn't about 'hacking' in the traditional sense; it's about quality control and reliability under stress. A model might be secure (un-hackable) but still fail because it is unreliable or hallucinates wildly.

Focus: 'Is it reliable? Does it fail gracefully?'
Core Functions:
- Validity & Accuracy: Does the AI actually answer the question asked, or does it confidently invent falsehoods?
- Resilience: If you input garbage data, adversarial noise, or out-of-distribution queries, does the system crash or handle it safely?
Genbounty Map: These findings fall under Core Safety Characteristics. They are crucial for enterprise clients who value stability over creativity.

3. The Values Layer: OECD Principles

The Human’s View (Fairness & Impact)

This layer assesses the impact on people and society. It is the foundation for global regulations like the EU AI Act. Vulnerabilities here are not 'bugs' in code, but 'bugs' in the model's worldview or alignment.

Focus: 'Is it fair? Is it legal?'
Key Principles:
- Non-Discrimination: Does the AI treat different user demographics equally? (e.g., denying loans based on race or gender).
- Transparency: Does the AI admit it is a machine? Can it explain why it made a decision?
Genbounty Map: These findings fall under Real-World Consumer Harms. These are critical for compliance-focused audits.

4. The Governance Wrapper: ISO 42001

The Regulator’s View

ISO/IEC 42001 wraps all the above into a certifiable management system. By finding a bug (OWASP), showing how it affects reliability (NIST), and documenting the societal impact (OECD), you help Genbounty clients achieve Grade A Certification.