--- url: "https://www.xcademia.com/news/google-details-how-it-tracks-ai-threats-and-builds-ai-defenses" title: Google Details How It Tracks AI Threats and Builds AI Defenses description: "Google outlines how AI is changing cyber threats and explains its approach using threat intelligence, multi-model AI, security graphs and automation." publishedAt: "2026-09-17T09:01:27.49+00:00" updatedAt: "2026-09-17T09:01:37.593083+00:00" type: news category: "cloud-security" source_name: Google Cloud Blog source_url: "https://cloud.google.com/blog/products/identity-security/cloud-ciso-perspectives-how-google-monitors-ai-threats-advances-ai-defenses" tags: - "#GoogleCloud" - "#Cybersecurity" - "#AISecurity" - "#ThreatIntelligence" - "#AIThreats" - "#CloudSecurity" - "#GoogleThreatIntelligence" - "#CyberDefense" --- # Google Details How It Tracks AI Threats and Builds AI Defenses > Google Cloud says AI is changing cybersecurity across software development, attack surfaces and threat capabilities. Its latest CISO perspective outlines how Google monitors these shifts and applies AI, threat intelligence and automation to defense. Source: **Google Cloud Blog** · 17 September 2026 Artificial intelligence is changing both sides of the cybersecurity equation. Attackers are using AI to accelerate activities ranging from software supply chain attacks to credential theft and automated intrusion workflows. At the same time, security teams are using AI to analyze threats, identify vulnerabilities and automate defensive actions. In its latest **Cloud CISO Perspectives** post, Google Cloud outlines how it is monitoring these developments and building defenses around them. Sandra Joyce, VP of Google Threat Intelligence, argues that security strategies around AI need to be grounded in observed threat activity rather than speculation. Google's position is shaped by its role across both AI development and cybersecurity operations. The company says this gives it visibility into how AI systems are built as well as how attackers are attempting to exploit them. Google Threat Intelligence Group, or GTIG, has identified **three structural shifts** in the threat landscape: 1. AI is changing how software is built. 2. AI is expanding the attack surface. 3. AI is increasing threat capabilities. The CISO perspective focuses on how Google is responding to each of these changes. ## 1. AI Is Changing How Software Is Built AI-assisted development is increasing the speed at which developers can write and deploy software. Google says autonomous agents and AI workflows are pushing code into production at increasingly high speed. While this can accelerate development, it also creates new security considerations for organizations trying to maintain oversight of rapidly changing software environments. One concern highlighted by Google is the possibility of attackers contaminating upstream software packages that AI coding assistants may recommend or trust. GTIG says malicious contamination of AI-assisted coding practices has contributed to the growth of large-scale open-source software supply chain compromises observed during 2025 and early 2026. The company's response is to place security controls directly into AI-assisted development workflows. Google describes this concept as a real-time **"spellcheck for cybersecurity."** The idea is straightforward: rather than requiring developers to leave their development environment to investigate every security issue, security controls can operate within the editor and agent workflow. Google says such controls could identify issues including poisoned packages, malicious prompts and misconfigured toolkits while development is taking place. ** Why Editor-Level Security Is Not Enough** Google also argues that securing the development environment alone does not provide complete visibility. A code editor does not necessarily have visibility into cloud configurations. A delivery pipeline may not have complete context about runtime exposure. Production security teams may also struggle to connect an active issue back to the underlying code or infrastructure configuration that created it. Google therefore advocates an integrated **code-to-cloud** approach in which code is continuously evaluated against the environment in which it will eventually operate. The company points to platforms such as Wiz Code as an example of this approach. ## Google Warns Against Single-Model Security Another issue highlighted in the CISO perspective is the use of a single AI model for security analysis. Google says relying on one frontier model can create what it describes as a security monoculture. The reasoning is that no individual AI model will necessarily identify every vulnerability, while attackers can deliberately test inputs against the weaknesses or safety mechanisms of particular models. Google says it therefore uses a **multi-model approach**. Its approach combines multiple foundation models, including Gemini as well as commercial and open-source models. According to Google, these models can be orchestrated to cross-validate findings, reduce false positives, remediate code and identify complex logic flaws that a single model may miss. The company presents model diversity as one component of its strategy for reducing dependence on any single AI system. ### ![info-1](https://0a515t3ure77wbvx.public.blob.vercel-storage.com/articles/1789635462465-info-1--152-.webp) ## 2. AI Is Expanding the Attack Surface Google says securing AI-assisted software development is only one part of the challenge. AI workloads themselves are becoming targets. The company highlights two areas in particular: **LLMJacking** and attacks targeting AI data and access. **LLMJacking Targets AI Compute** Google describes **LLMJacking** as attacks in which cybercriminals and state-sponsored groups target access to GPU resources so they can run their own AI models and agentic workflows. The post references an intrusion investigated by Mandiant in April. According to Google, the threat actor gained initial access to a victim's cloud environment through an exposed personal access token. The actor then used the access to deploy unauthorized AI infrastructure and scale high-performance computing resources. In this case, the victim was left responsible for the associated infrastructure costs. The example illustrates how cloud resources used for AI can become targets themselves rather than simply serving as infrastructure for legitimate workloads. ## AI Data and Credentials Are Becoming Security Targets Google also says attackers are increasingly targeting AI-specific information. This includes: - Custom prompts - Agent instructions - Fine-tuned models - Source code - AI skills - Related research Google reports that Mandiant investigated multiple data theft and extortion operations during **Q2 2026** involving proprietary AI data. The company also says demand for AI account credentials has increased in underground marketplaces. These developments expand the definition of sensitive AI assets beyond the models themselves. For organizations, prompts, instructions, model configurations and related development materials can also form part of the AI security boundary. ## Google Says AI Security Needs a Connected View Google argues that AI security should not be managed through disconnected security controls. Agent access policies, AI software inventories, shadow AI and runtime security can be closely connected. The company advocates a **unified and dynamic graph** connecting: **Code** **Models** **Data lineage** **Runtime identities** The objective is to provide a single contextual view of how different elements of an AI environment relate to each other. Google says the **Wiz Security Graph** provides the contextual foundation for its Google AI Threat Defense approach. The company describes Google AI Threat Defense as a broader security framework combining Gemini and other frontier models with Wiz's contextual risk prioritization, CodeMender's code remediation capabilities and Mandiant's threat intelligence and incident response expertise. Google says this context also feeds into **Google Security Operations**, where security teams can identify and prioritize attack paths. ### ![info-2](https://0a515t3ure77wbvx.public.blob.vercel-storage.com/articles/1789635480729-info-2--134-.webp) ## 3. AI Is Increasing the Speed of Threat Operations Google says attackers are moving beyond using AI simply to generate text or code. The company is observing increasingly automated, multi-agent attack workflows. One Mandiant investigation described in the post involved a financially motivated actor that compromised cloud infrastructure and deployed an autonomous agent framework. According to Google, the actor used an AI coding chatbot, a prompt and agent instructions to plan, build and execute a mass credential-harvesting campaign in **less than six hours**. Google also describes GTIG observations involving a PRC-nexus espionage group experimenting with a tool called **CC Switch**. According to the company, the tool was used to move between accounts and AI models, including Claude, Codex and Gemini, selecting models for different tasks. The underlying attack techniques were not necessarily new, Google says. The change was that AI could connect previously separate activities into a more automated workflow. These examples are Google's and Mandiant's reported observations, rather than independent measurements of the broader threat landscape. ## Defenders Have a Different Advantage: Context Google's argument is that attackers may have AI-enabled automation, but defenders possess something attackers generally lack: **deep internal context**. An organization knows its own: - Code - Cloud configurations - User identities - Deployment environments - Internal architecture Google says feeding this context into security models can help defenders connect signals that would otherwise appear separately. This forms the basis for the company's argument that AI can be used not simply to generate security alerts, but to continuously investigate and prioritize threats. Google says the goal is to move away from manually processing large numbers of static alerts toward machine-speed security operations. ## From Manual Response to Continuous Defense Google says autonomous security agents can be used to continuously monitor, investigate, prioritize and remediate attacks. This approach depends on combining AI models with organizational context and frontline threat intelligence. Rather than treating each alert as an isolated event, the system can use relationships between code, infrastructure, identities and runtime environments to establish broader context. The company describes this as a shift from **human-scale incident response** toward **machine-speed security operations**. However, Google does not provide a quantified comparison between these approaches in the announcement. ### ![info-3](https://0a515t3ure77wbvx.public.blob.vercel-storage.com/articles/1789635495344-ino-3--2-.webp) ## How Google Says It Is Defending the Ecosystem Google outlines four major areas where it says it is applying its security and AI capabilities. **Disabling Malicious Infrastructure** Google says projects, accounts and assets associated with known bad actors can be proactively disabled when its tools are being used to facilitate attacks. ## Hardening AI Models and Classifiers Google describes a continuous feedback loop in which threat intelligence is fed back into AI model and product development. The stated objective is to improve the ability of models to recognize and refuse malicious requests. ## Automating Vulnerability Discovery and Remediation Google says it is moving toward AI-driven vulnerability hunting and automated remediation. The company points to **CodeMender** as an example, describing it as a tool that can automatically fix critical vulnerabilities in code. ## Developing Specialized AI Defenses Google DeepMind teams are also working on specialized defenses for generative AI. Google says these efforts include active monitoring across its ecosystem to identify misuse in real time. ## A Continuous Security Model for the AI Era The overall approach described by Google connects threat intelligence, AI models, security context and automated response. The company argues that AI-driven attacks require defenses capable of operating at comparable speed. That does not mean eliminating human involvement. The source itself references both human-on-the-loop and human-in-the-loop approaches elsewhere in Google's security strategy. Instead, the emphasis is on using automation and AI to process large volumes of security information while incorporating organizational context and threat intelligence into the process. For security leaders, the central issue is therefore broader than simply adopting an AI security tool. It involves understanding how AI is entering software development, where AI workloads and data are exposed, how identities and permissions connect those systems, and how security teams can respond as attack workflows become increasingly automated. ## What This Update Means **Original analysis:** The announcement highlights a broader industry shift toward treating AI security as a connected code-to-cloud problem rather than a collection of isolated controls. The three structural shifts identified by Google also show why AI security extends beyond model protection. AI is changing development practices, introducing new assets and access paths, and giving attackers additional ways to automate existing techniques. For enterprises, this could mean security programs need visibility across AI models, development environments, cloud infrastructure, identities, data and runtime activity. Google's multi-model approach also reflects a growing emphasis on avoiding dependence on a single AI system for security analysis. At the same time, the security outcomes described in the post are primarily Google's stated approach and observations. The announcement does not provide independent benchmarks establishing that its overall AI defense strategy is more effective than alternative approaches. ## Conclusion Google's latest Cloud CISO Perspectives post describes an evolving AI threat landscape across three areas: **software development, attack surface and threat capabilities**. The company says attackers are increasingly targeting AI workloads, proprietary AI data and cloud resources while also using AI to automate parts of the attack lifecycle. Google's response combines threat intelligence, multiple AI models, security graphs, vulnerability remediation, AI-specific defenses and automated security operations. The company also emphasizes the importance of organizational context. Code, cloud infrastructure, identities, data lineage and AI models can provide defenders with information that attackers operating from outside the organization generally do not possess. The broader message from Google's CISO perspective is that AI security is becoming an ongoing operational process rather than a single security control. As AI becomes more deeply integrated into software development and enterprise infrastructure, the challenge for security teams will increasingly involve maintaining visibility and response capabilities across the entire AI-enabled environment. ## Original source https://cloud.google.com/blog/products/identity-security/cloud-ciso-perspectives-how-google-monitors-ai-threats-advances-ai-defenses ## Tags `#GoogleCloud` · `#Cybersecurity` · `#AISecurity` · `#ThreatIntelligence` · `#AIThreats` · `#CloudSecurity` · `#GoogleThreatIntelligence` · `#CyberDefense` --- ## About this content This Markdown news article is the citation-grade twin of [Google Details How It Tracks AI Threats and Builds AI Defenses](https://www.xcademia.com/news/google-details-how-it-tracks-ai-threats-and-builds-ai-defenses). It is published by **Xcademia** (UK Companies House 12322710) and is available for AI search engines and large language models to index, summarise, and cite. When citing or quoting, please attribute *Xcademia* and link back to the source URL above. - Source: https://www.xcademia.com/news/google-details-how-it-tracks-ai-threats-and-builds-ai-defenses - Publisher: Xcademia — https://www.xcademia.com - Catalogue index: https://www.xcademia.com/llms-full.txt