Attackers used AI agents after the PHP2Shell flaw, says Unit 42

Palo Alto Networks Unit 42 says attackers used AI agent software after they attacked a European government network through the PHP2Shell flaws. In less than two hours, the attacker got password hashes and secrets for multi-factor authentication. Unit 42 saw command output that an AI model can read, and Cloudflare says it saw millions of attacks after the public notice.

What Unit 42 saw

The attacker used commands with section headers in the output. Unit 42 says these headers make output that an AI model can read. The attacker then opened a tunnel into Active Directory, the directory of network accounts, and tried to get Domain Admin rights.

Scale of scanning

• A suspected North Korea campaign probes 216 hosts that have no patch, says CYFIRMA.
• Bitsight counts almost 37 million systems on the internet that use PHP or related software.
• The PHP Foundation warned of a third flaw. It needs an unusual setup.

What is not known

Unit 42 says the commands suggest an agent attacker. The source names no attacker for this attack. CYFIRMA connects its own campaign to the Lazarus Group with moderate confidence, but the source does not connect that campaign to the Unit 42 attack.

Insight: AI speeds up both the find and the attack

Researchers used the top AI models to find the PHP2Shell flaws, and the flaws leaked about nine hours before the planned notice. Unit 42 sees an agent attacker get secrets in less than two hours. It seems that AI shortens the time from flaw notice to deep attack, and a slow patch time helps the attacker. Count your patch time in hours, not weeks.