A software update goes live, and an automated system begins searching for the weakness it was designed to fix. It compares code, prepares an exploit and tests its way into vulnerable infrastructure without waiting for a human operator. Google Threat Intelligence Group and Mandiant say they have now observed this kind of autonomous activity in real intrusions. Their November 17 report describes a shift in cyber attacks that could leave businesses with far less time to secure exposed systems.

What the Company Did or Announced
Google published new findings documenting attackers' use of autonomous, multi-agent AI frameworks during active cyber operations. Its researchers identified tools including ARTEX, CAIRN, CyberStrikeAI and PentAGI, which can coordinate reconnaissance, vulnerability discovery and exploitation.
One particularly significant discovery involved an exposed server running a framework called exploit_agent. The system continuously examined software patches, identified potential vulnerabilities, generated exploits and executed them without human intervention. Evidence on the server pointed to interest in French media, entertainment and education websites.
Google also detailed its defensive response, including Google AI Threat Defense, which combines Gemini models, Wiz risk prioritization, CodeMender automated remediation and Mandiant intelligence.
The Numbers
The financial and operational consequences are substantial. During 2026, Mandiant investigated incidents in which criminals hijacked enterprise cloud resources to run AI services at victims' expense.
In some cases, fraudulent charges exceeded US$100,000 in a single day. The most severe cases accumulated US$1 million in resource abuse before remediation.

Another compromised environment experienced millions of AI API calls in approximately 24 hours. Separately, Google's vulnerability research found that monthly zero-day exploitation rose from a 2025 average of eight to 11 in 2026, before reaching 22 in August.
Those figures describe different risks: direct cloud billing exposure, rapid resource consumption and a growing volume of exploited software weaknesses. They are not a count of Canadian victims.
Why the Company Is Making This Move
The underlying problem is the shrinking interval between a vulnerability becoming visible and attackers exploiting it. Criminals can examine newly published patches to discover what changed, then automate the development of working attacks.
Google's research describes exploitation of PaperCut print-management vulnerabilities, followed by a campaign against hundreds of organizations. Researchers were able to produce working exploits within a day of disclosure, while subsequent malicious activity demonstrated how quickly attackers could move.
Google is responding by applying AI to defensive scanning, prioritization and patching. Its approach also includes model safeguards, account disruption and behavioural detection rules designed to identify malicious automation.
The company acknowledges a limitation in offensive AI: automated operations can produce recognizable traffic and forensic traces. That gives defenders opportunities to detect attacks even when execution is fast.
Impact on Employees, Customers, and the Industry
For Canadian organizations, the findings are relevant because the investigated techniques affect cloud services, software development environments and internet-facing infrastructure used internationally. Google's report identifies affected organizations across North America, although it does not provide a Canadian victim count.
Businesses face more than conventional data theft. Stolen cloud credentials can enable unauthorized AI subscriptions, unexpected computing bills and resale of access through underground services known as LLM Mixers.
Software teams face another exposure. Google documented suspected North Korean-linked activity involving modified Claude Code configuration files that could trigger a backdoor download when the coding assistant started. The investigated intrusions targeted cryptocurrency organizations in North America and Europe.
For employees, the findings also reinforce the risks associated with fraudulent recruiter profiles and deceptive meeting platforms used to obtain initial access.
What Comes Next
Google expects adversaries to continue adopting autonomous frameworks because they reduce the expertise and time required to conduct attacks. Its researchers also anticipate improvements in stealth as the underlying models and tools develop.
Defenders have practical priorities: reduce the time between patch availability and deployment, protect cloud credentials and session tokens, monitor unusual AI service consumption, and review automated development-tool configurations.
Google says it is strengthening model safeguards, deploying specialized detection capabilities and disabling infrastructure associated with identified misuse. Its full threat intelligence report provides further technical findings and mitigation references.
The central change is operational. Security teams are increasingly defending against systems that can discover weaknesses, adapt their methods and continue attacking with limited human direction.
Frequently Asked Questions
Can AI cyber attacks happen without a human controlling every step?
Yes. Google identified frameworks capable of discovering vulnerabilities, preparing exploits and executing attacks without continuous human intervention.
How do AI-powered attacks affect Canadian businesses?
Organizations using vulnerable software, cloud infrastructure or automated coding tools may face faster exploitation, credential theft and unauthorized computing charges.
What is LLMJacking?
LLMJacking involves criminals abusing compromised credentials or cloud resources to access AI services, sometimes producing large unexpected bills.
How quickly can attackers exploit newly disclosed vulnerabilities?
Google documented cases in which working exploits were developed within a day of disclosure, significantly reducing the time available for remediation.
What should organizations do to reduce their exposure?
Prioritize prompt security updates, strengthen credential protection, monitor cloud usage and review configurations that allow AI development tools to execute commands automatically.