AI About-Face: 'Mantis' Turns LLM Attackers Into Prey

Researchers from George Mason University present Mantis, a defensive system that embeds prompt-injection payloads into decoy services to misdirect or seize control of LLM-based attacking agents, achieving over 95% success in both passive and active countermeasures. The work details how direct and indirect prompt injection can exploit the iterative decision loops of automated penetration-testing AIs, even invisibly to human operators via ANSI-hiding, and argues that prompt-injection remains a hard-to-patch weakness, positioning human oversight as the only near-term mitigation.