Google launched AI Threat Defense this week, a platform it describes as an always-on security system for enterprises. The product brings together four assets under one workflow: Wiz for mapping what is exposed across a company's cloud infrastructure, Mandiant for real-world threat intelligence and incident response, CodeMender for automatically writing and testing code fixes, and Gemini for reasoning about what actually matters.
The context for this launch is not marketing. The numbers behind it are stark. CrowdStrike's most recent threat report recorded a fastest-ever criminal breakout time of 27 seconds, and an 89% increase in attacks by AI-enabled adversaries. The time between an initial breach and the next stage of an attack has dropped from eight hours to 22 seconds. Phishing attacks powered by AI rose 46% according to Microsoft's own data. The underlying point is simple: attacks now move faster than any human-managed security process.
Google's response is to automate the entire cycle. The platform first builds a live map of everything exposed, then runs multiple AI models against it to find real exploitable paths rather than generating a long list of theoretical warnings. From there, it automatically generates fixes, tests them before they go live, and tracks what was patched and when. Mandiant's role is to add real-world judgment, routing critical issues and helping teams handle sudden surges without overwhelming their engineers.
The Wiz piece is significant. Google paid $32 billion to close that acquisition in March, its largest deal ever. Wiz was already earning over $1 billion annually in revenue before the deal closed, and it works across AWS, Azure, and Oracle Cloud, not just Google. That matters because most large organizations are not running on a single cloud. Security problems do not stay inside one provider's environment.
CodeMender is worth understanding separately. It is an AI agent, originally built by Google DeepMind, that sits inside the tools developers already use daily and proposes fixes as they write code. The goal is to shrink the time between discovering a vulnerability and deploying a verified fix from weeks to minutes. Before any patch goes live, the platform runs automated tests. After it does, every fix is tagged with which AI model generated it and when, creating an audit trail.
Now the part that complicates the story. While Google is selling the idea of machine-speed defense, independent security researchers at Aikido found that when a Google Cloud access key is stolen, compromised, and then deleted by the owner, the key continues working for up to 23 minutes. The median window in their testing was 16 minutes. During that window, attackers can pull files and cached conversations from Gemini, access database services, and run up charges. Several developers were hit with bills exceeding $10,000 in roughly 30 minutes after attackers exploited this exact window.
Google classified the finding as "won't fix," describing it as a known property of its infrastructure. The researchers pointed out that Google's own newer key formats revoke in about a minute, and its service account credentials revoke in about five seconds. Both run at the same scale. The 23-minute gap for older keys is a matter of priorities, not engineering limits.
This does not make the advice wrong. The underlying shift, that manual security processes cannot match the speed of automated attacks, is correct and important. The global market for AI in cybersecurity is expected to grow from roughly $30 billion today to over $146 billion by 2034. Organizations using AI-powered security identify breaches significantly faster and reduce breach costs by close to $2 million per incident on average.
What it does mean is that any organization evaluating this kind of platform should ask hard questions about what the provider is and is not protecting in their own infrastructure. The pitch is to trust an automated defense system. That trust has to be earned by behavior, not just by announcements.