The Ransomware Crew Had 50 Attack Techniques Ready Before Lunch
A ransomware operator used artificial intelligence to compress more than 50 attack techniques into less than ten hours, according to new incident research from Palo Alto Networks Unit 42. That is the kind of intrusion sequence that has traditionally taken a human team days or even weeks to coordinate.
The important part is not that a chatbot wrote a phishing email. The attacker used frontier AI models and task-oriented agents after gaining initial access, then let those systems map services, search code, collect credentials and pursue several routes through the victim environment at once.
The intrusion moved at machine speed
Unit 42 mapped the activity to more than 50 MITRE ATT&CK techniques. The campaign unfolded in under ten hours, compared with an intrusion timeline that the researchers said could normally stretch to roughly two weeks.
The agents enumerated microservices, scraped repositories and located secrets that exposed root credentials. They then attempted to abuse CI/CD workflows and modify infrastructure-as-code. One Terraform backdoor attempt failed because branch protections blocked the change, a useful reminder that ordinary engineering controls can still stop an AI-assisted attacker.
The AI became part of the attack infrastructure
The operator did not rely on one long conversation with a model. The activity showed multiple agents working in parallel, exchanging structured Markdown and leaving behind scripts designed to keep the campaign moving. After cloud credentials were stolen, the attacker also used the victim’s own AI endpoints as post-compromise infrastructure.
That detail changes the defensive problem. Requests to an approved AI service can look like normal business traffic. If the attacker uses the victim’s account, endpoint and billing relationship, defenders may see authorised infrastructure performing malicious work.
It even wrote the handover notes
The campaign left an 80-page technical audit of the compromised environment. What looks like helpful documentation for a legitimate security team can also become a persistence aid for criminals. A new operator can inherit a detailed map of identities, services, weaknesses and unfinished tasks without repeating the discovery phase.
Unit 42 stressed that a human still set the goals and made consequential decisions. The agents adapted, replanned and executed many individual tasks, but they did not independently invent the criminal objective. That distinction matters when assessing current capabilities without exaggerating them.
What defenders should watch for
Traditional alerts remain important, but the rhythm of the activity may be just as revealing. Parallel calls to large language models, rapid creation of custom scripts, structured notes passed between sessions and sudden use of internal AI endpoints can expose an agentic workflow.
Security teams should place strong controls around code repositories, secrets managers, CI/CD runners and infrastructure-as-code. Protected branches, scoped credentials, approval gates and workload identity boundaries are no longer just development hygiene. They are brakes that can slow an intrusion moving at machine speed.
The lesson is uncomfortable but practical. AI did not remove the need for an initial foothold or human intent. It removed much of the waiting between one successful step and the next.
