Key Takeaways
- Upwind’s security research team, led by Uriel Yochpaz, drives product innovation by thinking like attackers and uncovering real‑world threat techniques.
- The team (~30 researchers) is split into Platform, Threat & Runtime, AI, and Offensive groups, each focusing on cloud, AI infrastructure, threat intelligence, AI‑enabled defenses, and proactive vulnerability hunting.
- Research follows a disciplined process: mapping legitimate use, identifying attack surfaces, adopting an adversarial mindset, and testing in isolated lab environments with proxy, AI‑powered, and disassembly tools.
- Prioritization weighs customer relevance, broad adoption, potential severity, and alignment with Upwind’s cloud and AI domains.
- A flagship discovery showed how the AWS Data Perimeter could be bypassed using Amazon Bedrock AgentCore, creating a bidirectional command‑and‑control (C2) channel that abused legitimate AWS behavior.
- Responsible disclosure prompted AWS to fix the infiltration channel; the exfiltration path remains under review as standard OIDC behavior.
- Yochpaz’s “Moby Dick” is a cross‑tenant cloud vulnerability that would break provider isolation—a scenario with catastrophic impact.
- While inter‑team camaraderie fuels internal collaboration, healthy competition with other companies pushes researchers to raise the bar, echoing the Talmudic idea that scholarly envy increases wisdom.
- AI amplifies researcher productivity exponentially, but human expertise remains essential to steer AI tools toward relevant findings and to tackle emerging AI‑specific threats such as prompt injection and MCP‑related risks.
- The future of security research lies in the symbiosis of human intuition and AI speed, enabling teams like Upwind’s to stay ahead of an ever‑accelerating threat landscape.
The Role of Research at Upwind
In many ways, a great deal of what the company builds starts with the research team,” says Uriel Yochpaz, Vulnerability Research Lead at Upwind. By continuously thinking like an adversary and uncovering new attack techniques, the team ensures the company stays ahead of the curve and maintains a finger on the pulse of the evolving threat landscape. This proactive stance translates directly into product capabilities that address real‑world risks before they are exploited in the wild.
Yochpaz’s Journey from IDF to Upwind
Yochpaz’s career in offensive security began in the IDF’s J6 & Cyber Defense Directorate, where he served as a Red Teamer before being promoted to Team Lead overseeing a group of security researchers. He later joined Unit 8200 as an Offensive Security Researcher, leading complex network operations and developing custom offensive tooling. After a stint at Upstream Security’s CTO office—where he uncovered vulnerabilities across client‑facing platforms and integrated GenAI capabilities into an API security product—he now leads the vulnerability research team at Upwind. His background includes multiple CVEs and zero‑day discoveries, an OSCP certification, a competitive cyber CTF win, and a B.Sc. in Computer Science from the Open University.
Composition and Specializations of the Research Department
Upwind’s Security Research Department comprises roughly 30 researchers distributed across four specialized groups. The Platform Research team concentrates on securing cloud and AI infrastructure, aiming to build the industry‑leading Cloud Security Posture Management (CSPM) solution. The Threat and Runtime Research department focuses on detecting and tracking adversaries while advancing threat intelligence knowledge. The AI Department creates innovations that leverage advanced AI models to revolutionize cybersecurity defenses. Finally, the Offensive Department designs offensive security tools to validate findings from other groups and conducts proactive vulnerability research to uncover flaws before threat actors can exploit them.
Research Process and Methodology
When a target—be it a specific technology or platform—is selected for research, the team first seeks to understand its legitimate, intended use and the full spectrum of functionality it offers. Next, they map the attack surface and analyze the existing threat model. Adopting an attacker’s mindset for each surface, they look for ways to abuse legitimate capabilities to achieve outcomes that should not be possible through normal use. This adversarial perspective is the cornerstone of discovering novel weaknesses that defenders might overlook.
Topic Prioritization and Laboratory Setup
Research topics are prioritized based on several criteria: relevance to customers (so findings deliver direct value to them and to Upwind’s product), broad real‑world adoption (maximizing impact), potential severity (focusing on flaws that could cause significant or dangerous consequences), and alignment with Upwind’s core cloud and AI domains. Once a topic is chosen, the team establishes a dedicated lab environment that allows secure, isolated experimentation. Their toolkit includes proxy tools for intercepting and manipulating network traffic, AI‑powered utilities and MCP servers, and disassemblers for low‑level binary analysis. While everyone works on overlapping themes, each researcher cultivates a niche specialty, contributing deep expertise that elevates the collective effort.
How Research Drives Product and Trust
High‑quality, comprehensive research feeds the product team with the insights and conclusions needed to develop new capabilities and push the platform forward. By continuously thinking like an adversary and uncovering fresh attack techniques, the research team ensures Upwind stays ahead of emerging threats and maintains a finger on the pulse of the threat landscape. Beyond product impact, the discovery of complex, sophisticated vulnerabilities positions the company as a deeply technically credible organization—one that genuinely understands the threat environment at an expert level. This reputation fosters trust with customers and within the broader security community, reinforcing Upwind’s market position.
Significant Finding: AWS Data Perimeter Bypass
One of the team’s most notable discoveries was a novel attack technique that bypasses the AWS Data Perimeter—a layered defense model enforcing trusted identity, trusted resources, and expected network origin for every API call—using Amazon Bedrock AgentCore, AWS’s managed service for deploying AI agents. The research demonstrated that an attacker with a single code‑execution foothold inside a protected environment could establish a full bidirectional Command‑and‑Control (C2) channel by abusing two legitimate behaviors of the AgentCore service.
Mechanics of the Bidirectional Command‑and‑Control Channel
The first channel enabled data exfiltration: when an AgentCore Runtime is configured with a JWT discovery URL, the service issues a server‑side HTTP request to validate that URL, originating from AWS infrastructure itself rather than from the customer’s VPC. By encoding sensitive data into the URL path, an attacker could exfiltrate information each time the runtime updated. The second channel facilitated command infiltration: an unauthenticated API endpoint—GetRuntimeProtectedResourceMetadata—remained accessible through VPC endpoints even under strict deny‑all policies and produced no CloudTrail logs. Together, these two channels formed a complete bidirectional C2 using only legitimate AWS service functionality, illustrating how trusted services can be repurposed for malicious purposes when subtle misconfigurations or design nuances are overlooked.
Responsible Disclosure and AWS’s Fix
The findings were presented at fwd:cloudsec North America 2026. Following responsible disclosure, AWS deployed a fix for the infiltration channel across all production regions. The exfiltration channel, which AWS considers standard OIDC behavior, remains open as the provider evaluates additional defense‑in‑depth improvements. This outcome underscores the importance of coordinated vulnerability disclosure and the ongoing dialogue between security researchers and cloud providers to harden shared infrastructures.
Aspirational Threats and Industry Collaboration
When asked about his personal “Moby Dick,” Yochpaz described a cross‑tenant vulnerability in a cloud environment—a flaw that could compromise every customer of the affected provider by breaking the fundamental isolation that underpins the multi‑tenant model. A close second is unauthenticated remote code execution (RCE) with no prior authentication or pre‑existing foothold, a scenario representing one of the most critical findings imaginable. Regarding competition, he notes a strong sense of camaraderie within Upwind’s teams, where one researcher’s success lifts the whole company. Externally, there is considerable knowledge sharing and humility; researchers openly discuss findings to enrich collective wisdom. He cites the Talmudic maxim “the envy of scholars increases wisdom” (kinat sofrim tarbeh chochma) to illustrate how healthy rivalry drives everyone to greater heights.
The Evolving Role of Humans Amid AI‑Powered Research
Yochpaz views the future of human security researchers through two lenses. First, AI exponentially amplifies researcher productivity: problems that once required days of effort can now be solved in seconds or minutes, enabling rapid vulnerability discovery and weaponization. Nevertheless, an experienced researcher remains essential to steer AI tools, filter out irrelevant noise, and pinpoint precisely what is relevant to the investigation. Second, AI introduces an entirely new domain demanding dedicated security attention—from prompt injection attacks to the security challenges of Model Context Protocol (MCP) and other AI‑specific interfaces. Thus, while AI reshapes the researcher’s toolkit, human judgment, creativity, and deep contextual understanding remain irreplaceable.
Looking Ahead: Staying Ahead of the Threat Landscape
Upwind’s research team exemplifies how a disciplined, adversarial mindset coupled with cutting‑edge AI assistance can yield high‑impact discoveries that shape product development and bolster industry trust. As cloud and AI ecosystems continue to expand and intertwine, the need for proactive, specialized research will only grow. By maintaining a clear focus on customer relevance, severity, and domain alignment, and by fostering both internal collaboration and healthy external competition, Upwind’s researchers are well positioned to anticipate and neutralize the next generation of cyber threats—ensuring that security stays one step ahead in the real‑time era.