Google Cloud Offers $250,000 for Critical Cloud Architecture Vulnerabilities
The new $250,000 Google Cloud vulnerability reward is a bet on human judgment. Automated tools can produce more vulnerability reports, but Google says the number of valid submissions to its Cloud Vulnerability Reward Program has held steady. The prize is aimed at findings that require researchers to reason across cloud architecture, especially flaws that could break tenant isolation or expose control across services.
That distinction matters. More reports don’t automatically mean more useful discoveries, and faster discovery can leave defenders less time to respond. The new tier rewards a narrow, high-impact result: proving full administrative control over an arbitrary Google Cloud organization or project, without prior access, stolen credentials, user interaction, or a specially prepared target.
Chris Betz, Google Cloud’s CISO, discusses the program with Michael Cote, its vulnerability rewards lead, in the second October 2026 edition of Cloud CISO Perspectives. Their conversation ranges from safe validation of cloud takeovers to the risks of rushing vulnerability disclosure as AI tools make research more accessible.
The newsletter is also published on the Google Cloud blog. Readers can subscribe to receive the email edition.
Why Google created a $250,000 Root-in-Cloud reward
By Chris Betz, CISO, Google Cloud
Google Cloud’s Vulnerability Reward Program (VRP) works with security researchers around the world to uncover serious software vulnerabilities. The new top-tier reward reflects a concern that’s easy to miss amid the surge in automated discovery: Google says it needs more high-level research led by people who can reason about complex systems, with agentic tools as an aid.
The $250,000 Root-in-Cloud reward targets architectural flaws with potentially broad consequences. Finding a tenant-isolation bypass or a subtle logic weakness spanning multiple services takes more than scanning code; it calls for systems analysis and human insight, often alongside automated agents.
To qualify, a researcher must show full administrative control—equivalent to the roles/owner role or complete IAM control—over an arbitrary Google Cloud organization or project. The target must require no prior access or special setup. An exploit cannot rely on a foothold in the victim’s project, leaked, phished, or stolen credentials, user interaction, or social engineering.
Proving access without putting customers at risk
Google has published technical criteria for validating root-in-cloud access, intended to make testing repeatable without touching production customer environments. Reports showing remote code execution in Borg, Google’s underlying production infrastructure, or achieving root-in-prod will also be reviewed by the Cloud VRP panel for possible Root-in-Cloud eligibility.
There are strict limits. Researchers who achieve code execution in Borg must stop probing immediately, because further activity could affect service availability or expose data across tenants. Escalation, extracting production credentials, and lateral movement are prohibited and can disqualify a researcher. The panel will assess the initial entry point, execution context, and attack surface to decide whether that foothold could enable arbitrary tenant or project takeover—and determine the reward.
The new tier arrives as security research changes. At Google Security’s annual Escal8 conference in Singapore the previous week, Betz spoke with Michael Cote about the program and the wider effects of agentic tools. Their exchange has been lightly edited.
Chris Betz: You’re at Escal8 in Singapore announcing a $250,000 Root-in-Cloud tier. Why launch it now, and what research do you want it to encourage?
Michael Cote: It recognizes how security research has changed over the past year. We value strong reports from anyone, and we’re seeing that human creativity combined with agentic tools can produce more than either can alone.
Scanners can flood programs with low-impact findings. Meanwhile, first-time researchers using their intuition with agents are finding more interesting bugs, on average, than novices were two years ago. We want the new tier to push researchers toward complex, consequential discoveries—such as tenant-isolation bypasses and cross-service logic flaws—that demand architectural insight. Those findings help protect customers.
Betz: Google Threat Intelligence Group (GTIG) recently reported that monthly CVE disclosures have doubled as agentic tools mature. What has automation changed in the reports coming into Cloud VRP?
Cote: It has greatly increased the volume of submissions, though the number of valid vulnerability reports has stayed stable over the past 18 months. More people can now try vulnerability research, and experienced researchers at Escal8 said the barrier to entry is falling quickly.
That brings more newcomers with interesting findings and puts pressure on seasoned researchers to work faster and find higher-impact bugs. Experience still makes a difference.
Betz: You’ve worked with the research community for more than 20 years and value its commitment to making the world safer. How is that community changing as agentic tools spread?
Cote: That shared commitment remains central, but pressure to move faster deserves attention. GTIG’s research found the sharpest increase in exploited vulnerabilities involves n-days rather than zero-days.
Open source accounts have suggested that threat actors use large language models to turn newly public vulnerability details into working exploits quickly. When triage teams are already dealing with more reports, publishing an unmitigated vulnerability before the standard 90-day coordinated disclosure window can hand attackers both the information and the time they need to target customers.
The growing submission volume may also encourage researchers to publish before they have fully checked a finding. As LLMs make it easier to find bugs in source-available software, researchers may fear that someone else will publish first. Even established security companies can feel that competitive pull. Disclosing a vulnerability before a fix or mitigation is ready puts users at risk.
Project Zero helped establish the industry’s 90-plus-30-day responsible disclosure policy. Autonomous discovery and exploitation make coordinated disclosure more important. A company may get brief attention by publishing first, but customers are unlikely to trust organizations that put them and the wider ecosystem at risk.
Betz: How can researchers validate a finding for a $250,000 bounty without endangering customer projects or service availability?
Cote: A large reward needs clear rules for safe testing. Researchers should never test exploits against real customer projects. Instead, we created a deterministic proof-of-root process based on a designated service account, allowing researchers to cryptographically demonstrate unauthorized access to that identity.
We’ve also updated the program rules with firm boundaries for underlying production infrastructure. If researchers are unsure whether a test crosses a line, they should contact us. We’re happy to help them demonstrate the value of their work without risking a disqualification.
Betz: What should enterprise CISOs running vulnerability programs and working with outside researchers take from this shift toward AI-assisted research?
Cote: First, prepare triage processes for submissions generated with agents. Teams need to distinguish plausible-sounding hallucinations from real security issues, so valid reports don’t get buried in a backlog.
Second, shorten remediation timelines after a vulnerability becomes public. Autonomous tools reduce the time between disclosure and a working exploit, so defenses have to catch up faster.
Finally, expect security partners and vendors to follow coordinated disclosure standards. The industry should normalize agentic vulnerability research while keeping user safety at its center.
Google Cloud’s VRP rules page includes endpoint specifications, JSON payloads, and command-line examples.
More from Google Cloud security
Recent updates from Google’s security teams include:
- How Google Does It: Using agents to help find, track, and fix vulnerabilities: Google describes using agentic systems alongside existing security foundations as threat actors find software flaws faster. Read more.
- How to use agents to stop insider threats: Guidance on ensuring enterprise agents neither become insider threats nor help people act as one. Read more.
- Responding to core PHP vulnerabilities CVE-2026-109618 and CVE-2026-10961: Recommendations for reducing remote code execution risk from two critical vulnerabilities in the PHP runtime. Read more.
- Wiz Sensor for GKE: Turnkey threat detection for Kubernetes: Google Kubernetes Engine’s Wiz Runtime Sensor add-on is in preview, with tools for deployment, lifecycle, credential security, and governance. Read more.
- Scale agents securely with IAM Helper Agent skills and MCP tools: Google introduces IAM Helper, agent-facing components designed to automate end-to-end IAM workflows. Read more.
- Introducing new scalable perimeter options with VPC Service Controls: Google announces changes intended to make VPC Service Controls perimeters easier to create, maintain, and scale. Read more.
Security updates: privacy, supply chains and active threats
Two announcements point to the same pressure point: teams need to protect systems while responding faster to threats. Google says Confidential Computing is now generally available on G4 GPUs, bringing hardware-based security to AI Hypercomputer while supporting dynamic performance. Read more.
Wiz describes a supply chain pipeline designed to catch compromised packages as they’re published, before a CVE appears. That timing matters: waiting for a formal vulnerability notice can leave defenders behind. Read more.
For more security stories from Google Cloud, see what was published this month.
Threat intelligence: short attack windows, familiar exposure
The supply chain threat is getting harder to catch with routine scanning. Google Threat Intelligence Group (GTIG) and Mandiant report encountering a compromise in which malicious changes were active for just five to 10 minutes. Such brief windows can evade point-in-time triage and automated file scans. Read more about hardening code pipelines and CI/CD infrastructure.
There’s also a warning for organizations running BeyondTrust Remote Support appliances. During an incident response engagement, Mandiant found evidence that ARMORY CASTLE, a suspected PRC-nexus espionage actor, exploited a vulnerability in the appliances. Once a vulnerability becomes public, ARMORY CASTLE and other actors have historically moved to broad, opportunistic exploitation. Read more.
The broader advice is to rethink vulnerability response as discovery and exploitation speed up in the AI era. The recommendation is to move away from mass patching without prioritization and toward threat intelligence-driven triage, targeted edge defenses and automated, agentic remediation. The data doesn’t establish how quickly every organization can make that shift, but it makes the old patch-everything approach look increasingly blunt. Read more.
For an immediate case study, Mandiant Consulting and GTIG identified active, in-the-wild exploitation of a zero-day vulnerability in Citrix NetScaler ADC and NetScaler Gateway appliances. Their post covers containment and remediation guidance. Read the guidance.
Google Cloud has more threat intelligence stories published this month.
Listen: AI, vulnerability research and secure design
These episodes examine how AI is changing security work, from finding flaws to fixing them. In one, Jasika Bawa and Doug Turner of Chrome Security discuss how Google Chrome uses AI agents to identify and patch vulnerabilities autonomously, accelerating the browser’s update cadence. Listen to “Patching browsers with AI, agents, Rust, and your tabs”.
Wiz head of vulnerability research Nir Orfeld explains how his team uses multi-agent AI systems to find high-impact zero-day vulnerabilities in cloud infrastructure through “All about Project Atlas, Wiz’s AI vulnerability research”.
And Google principal security engineer Christoph Kern discusses secure-by-design practices for an enterprise operating at Google’s scale, including the challenge of code written by AI when reviewers have limited time. Listen to “How Google eliminates classes of vulnerabilities at scale”.
For another Google Cloud security discussion, watch now.
Sign up to receive Cloud CISO Perspectives twice a month in your inbox through the Google Cloud CISO newsletter. More Google Cloud security updates are due in a few weeks.