What is data poisoning and how would you defend against...

Data poisoning involves inserting malicious text to steer model outputs. For platforms like Wikipedia, the attack surface is the window between edits and rollbacks; mitigation involves diffing snapshots against a trusted live reference to flag transient, potentially malicious content. For web-scale data like Common Crawl, domain-level reputation signals are effective for excluding known spam or compromised sources.