Joey Melo Explains AI Jailbreaking and Data Poisoning

Summary

CrowdStrike researcher Joey Melo describes how AI red teams probe model guardrails through prompt and context manipulation, and test susceptibility to data poisoning. He says jailbreaks have become harder, but ongoing AI changes bring new risks.

Key points

• Melo describes AI red teaming as probing how prompts can bypass a model’s guardrails without changing its source code.
• His approach begins by testing a bot’s intended role and limits, then varying prompts and context to see how its responses change.
• Context manipulation can try to override a model’s existing rules by supplying new assumptions or instructions.
• Data poisoning tests whether prompts or external content ingested for training can influence a model’s later outputs.
• Melo says AI jailbreaks have become more difficult over the past two years, while continued changes to AI can introduce new risks.
• He says responsible disclosure and helping developers improve guardrails are central to his ethical red-team work.

People