Palo Alto Networks details 10-hour AI agent hack

Palo Alto Networks walked Semafor through a hack that played out over roughly 10 hours, starting with a human attacker who handed reconnaissance to an AI agent. That agent ran a broad internet scan and found a public API endpoint — an opening companies leave so outside software can talk to their systems. Piazza's point is that this kind of scanning is constant and hard to prevent, which is what made the entry point attractive.

Once inside, the intruder switched to internal reconnaissance, mapping what systems and software were running. The coding pipeline and code repositories stood out as targets. Sub-agents then combed those repositories and pulled out credentials — passwords and tokens developers had left written directly into the company's software. Those logins opened the system the company uses to build and deploy software, and from there the attackers took the keys to its cloud accounts, running the operation through the victim's own AI tools. Piazza described this as 'living off the land,' which makes the activity harder to detect.

His reading is that this playbook could spread, because AI systems are getting more access to sensitive company data. For security teams, the practical question is whether credentials are properly locked down; Piazza framed that discipline as what keeps a lot of low-grade attacks out. The outcome here hinged on the intrusion being spotted in time, and it was — the attacker demanded a ransom but was ejected before succeeding, though an AI agent left behind an 80-page technical report on the company's vulnerabilities.