Data-Void Exploits manipulates search-grounded AI via one attacker web page

Evidence record

Read the original source ↗

Claim-level citations (2)

• supportsSearch-grounded models reused attacker pages where independent evidence was scarce: The Data-Void Exploits paper reports a 56-topic controlled evaluation in which manipulated pages entered model answers and citations under scarce independent coverage.

  Abstract; data-void experiments; citation and reuse measures

• supportsSearch-grounded models reused attacker pages where independent evidence was scarce: The Data-Void Exploits controlled study found attacker content appeared in 31% of tested search-model responses across prompting, hosting-visibility and content-design conditions, without requiring control of more than one attacker page.

  Abstract; one-page attacker control and 31% aggregated retrieved/cited/used response outcome