AI-Assisted Attack Chain Compromises 25 Servers and a U.S. Government-Affiliated Network

Security firm Dream has detailed an attack, observed on October 25, 2026, that appears to have been assisted by artificial intelligence. It is attributed to a suspected China-nexus threat actor. The attacker combined the PHP2Shell vulnerabilities with a Linux kernel privilege escalation flaw, tracked as CVE-2026-46300, to gain full control of 25 servers across the U.S., Japan, Taiwan and South Korea. Three of those systems sat in a U.S. government-affiliated environment.

The affected devices ran specific versions of a large vendor's firewall software. The vendor was not named, but Dream says it notified the entity. According to the researchers, the exploit was a custom, non-public build tailored to the targeted firewall platform, not a copy of a known public exploit. It targets PHP's session.upload_progress feature on the firewall's management server, which runs PHP 8.0.30 with a high memory limit. After gaining control, the attacker went after Active Directory, and recovered files confirm credential extraction and directory enumeration.

The AI link comes from evidence found in the attack infrastructure. This included an installation of Anthropic's Claude Code, plus API keys for FOFA and for an AI-powered offensive security platform called aiscan, configured with large language models from DeepSeek and Kimi. Dream's artifacts suggest the attacker developed the exploit in a lab before gaining persistent access to a U.S. government organisation's internal Active Directory domain.