Threat actors exploited legitimate AI security tools at 90+ organizations using malicious prompt injection, with the next wave gaining write access to firewalls.