Meta Attacks for AI: Human and AI Reviewers Under Attack
Description
Currently, guardrails are implemented in generative AI systems, such as chat-based AI, to detect malicious users and to prevent harmful outputs through automated filtering and content moderation in human-in-the-loop AI systems that involve both automated and human review.
However, explicit attacks that target the guardrails themselves remain underexplored. In this work, I examine this problem and define the following concepts:
Meta Attacks / Infiltration for AI
- Guardrail saturation attacks (Saturation Attacks), including:
- AI-review-based guardrail saturation attacks
- Human-review-based guardrail saturation attacks
- Exposure enhancement via human and AI-based review systems (including AI/AGI/ASI)
- Attacks on human reviewers (i.e., content moderators) via guardrails
- Attacks on AI-based review systems via guardrails
Unlike jailbreak or prompt injection attacks that aim to extract prohibited outputs,
these Meta Attacks target the guardrail mechanisms themselves, including human and AI reviewers.
Related links and updates are available at: