Google Cloud Security has unveiled a unique architecture for an agent-based system that identified over 100 confirmed critical vulnerabilities within just two days. This internal framework combines multiple agents, Mandiant’s expertise, and human validation, providing Google with a foundation for offering comprehensive defense solutions—from code discovery to remediation and monitoring.
For the past ten months, Mandiant has utilized the Agentic Vulnerability Discovery Harness (AVDH) to analyze source code repositories. During an operation involving compromised enterprise repositories, the system sifted through tens of millions of lines of code to produce more than 100 confirmed critical findings in just 48 hours. According to Google Threat Intelligence, this initiative has already resulted in the attribution of twelve Common Vulnerabilities and Exposures (CVEs), including CVE-2026-13242 and CVE-2026-55803, with an additional dozen vulnerabilities still undergoing coordinated disclosure.
This large-scale vulnerability search also accelerates the discovery of exploitation paths. Consequently, security labs are enhancing their defensive capabilities while ensuring a high technical cost for potential attackers. Anthropic differentiates Fable 5 from Mythos 5 based on their safety mechanisms and access protocols. OpenAI restricts Daybreak to select partners, while Microsoft is currently testing MDASH with a limited number of clients. Meanwhile, Google distributes tasks across various agents, assesses exploitability, and submits findings to specialists, creating a formidable AI-driven cybersecurity defense that can protect models, monitor their usage, and deploy capabilities against adversaries.
AVDH Transforms the Model into a Comprehensive Investigation Framework
The architecture detailed by Google Threat Intelligence allocates analysis tasks among various specialized agents. The first step establishes the threat model and entry points, followed by additional agents that enrich context, monitor access controls, and hypothesize on data flows. Several validation agents then test each hypothesis using diverse configurations, while a synthesis agent consolidates their conclusions before a final review by a Mandiant expert.
AVDH addresses a significant challenge often encountered with direct queries to models: such models may deliver lengthy lists of plausible weaknesses without assessing their exploitability. AVDH enforces analytical pathways, cross-references multiple responses, and keeps humans in the loop when decisions could lead to disclosures or fixes. Mandiant enhances the models using rules derived from its incident response operations. The ultimate output relies as much on specialized knowledge and validation as on the model itself.
Currently, Google restricts AVDH to its internal analyses, providing no details on pricing, timelines, or false positive rates.
Mythos and Fable Expose Supply Chain Risks
The fallout from Mythos and Fable has highlighted supply chain vulnerabilities. On June 12, 2026, the U.S. Department of Commerce mandated Anthropic to suspend foreign nationals’ access to both models. As a result, Anthropic discontinued access for all clients, unable to verify users’ nationalities in real-time. The restrictions were lifted on June 30, with Fable returning to the global catalog while Mythos was initially reinstated for select approved U.S. organizations.
This directive disrupted services provided by Anthropic and distributed through AWS, Google Cloud, and Microsoft. Marketplaces do not mitigate risk when the same model can be withdrawn from the source. Cyber teams must evaluate continuity of access to frontier models, assess alternatives, and consider re-running their procedures with different models. Distribution mechanisms therefore become a critical product feature.
Google’s multi-model approach lessens this dependency by embracing alternative models. CodeMender allows users to select from different models based on cost, speed, or depth of analysis. However, this flexibility does not exempt Google from American regulatory decisions. The company remains subject to U.S. law, and certain capabilities of Gemini 3.5 Flash Cyber are reserved for a limited number of trusted governments and partners. The real advantage lies in the ability to replace components without the need to rebuild the entire defense chain.
Google Links Code to Business Risk
Google now possesses the essential building blocks to evolve AVDH into a continuous security service. Google AI Threat Defense integrates Gemini and other models, CodeMender for code verification and correction, Wiz for connecting vulnerabilities to exposed resources, and Mandiant for intelligence and incident response. Google Security Operations contributes detection, investigation, and response capabilities across production environments.
This system combines Wiz and CodeMender to address the new bottleneck faced by security teams. The agents can identify more vulnerabilities than developers can remediate, leading to a growing backlog of tickets and alert fatigue. Wiz provides contextual deployment information, identities, accessible data, and attack paths, while CodeMender executes proof-of-concept tests in a client-managed sandbox, generates a correction, and presents it in the form of code changes. Developers maintain the final approval before any integration.
Google’s acquisition of Wiz for $32 billion underscores its ambitious commercial aspirations. Finalized in March 2026, this acquisition promises seamless onboarding of code from cloud to execution across major cloud providers, though several integrations remain in progress. The CodeMender page indicates that product calls from Wiz are still forthcoming, and access to the specialized cyber model is restricted, with pricing details yet to be disclosed.
Meanwhile, Microsoft continues to restrict MDASH to a private preview phase. This tool orchestrates over 100 agents and has contributed to the discovery of sixteen vulnerabilities in Windows. OpenAI limits direct access to GPT-5.5-Cyber to vetted defenders and partners. Competition in the space is shifting towards procedures, exploitability verification, remediation, and seamless integration with existing tools, with the model evolving into a replaceable resource within a broader system.
Google AI Threat Defense stands to make significant strides by selling an integrated results chain rather than a direct equivalent to Mythos. Its competitive edge lies in the synergy among Gemini, Wiz, Mandiant, CodeMender, and its security operations. Companies would gain the ability to identify, prioritize, remediate, and monitor vulnerabilities effectively. However, the promise of this system hinges on the availability of its components, their multi-model interoperability, and evidence of effectiveness from clients. While the architecture is established, transitioning it into a cohesive service will be its next critical test.



