Vulnerabilities & CVEs

Pwn2Own Berlin: 47 Zero-Days Found, AI & Enterprise Systems

The recent Pwn2Own Berlin event wasn't just about bragging rights; it was a stark reminder of the vulnerabilities lurking in the enterprise AI stack. Almost $1.3 million was awarded for uncovering 47 zero-day flaws.

Cybersecurity researchers working at computers during Pwn2Own Berlin event

Key Takeaways

  • 47 zero-day vulnerabilities were discovered at Pwn2Own Berlin, targeting enterprise AI and development tools.
  • The Devcore team took home the largest prize ($505,000) by successfully chaining multiple vulnerabilities.
  • The event highlights AI databases and coding agents as significant new attack surfaces.
  • Researchers earned nearly $1.3 million in total for their findings, underscoring the value and prevalence of these undiscovered flaws.

Forget the headlines about abstract cybersecurity threats; what happened at Pwn2Own Berlin translates directly to your inbox, your cloud deployments, and the very code you’re building. When security researchers glean nearly $1.3 million by finding 47 zero-day vulnerabilities in enterprise AI databases, coding agents, and NVIDIA products, it means the tools we’re increasingly relying on are far from bulletproof. This isn’t just a technical footnote; it’s a warning shot for every organization pushing AI into production.

This year’s Pwn2Own Berlin, a three-day contest sponsored by Trend Micro’s Zero Day Initiative (ZDI), wasn’t your typical consumer-facing gadget hackathon. It was laser-focused on the enterprise backbone – the AI infrastructure that’s rapidly becoming the engine room of modern business. Teams went after AI databases like Chroma and Postgres pgvector, coding assistants such as Cursor and Claude Code, and even NVIDIA’s core offerings. The sheer number of zero-days – flaws unknown to vendors and therefore unpatched – is frankly staggering. And the $1.3 million payout? That’s a clear indicator of the high-stakes nature of these discoveries.

What’s especially telling is the specific emphasis on AI and coding agents. Dustin Childs, ZDI’s head of threat awareness, put it starkly: “At some point or another, we’ve probably all vibe coded something. There’s no shame in that, but how secure are the tools we use for vibe coding?” The implication is clear: the “vibe coding” — that rapid, iterative development often enabled by AI assistance — is now a prime attack vector. Imagine a malicious actor compromising a coding agent, subtly injecting backdoors into thousands of projects, or worse, exfiltrating sensitive code and proprietary data. It’s a chilling prospect that underscores an architectural shift: our development pipelines themselves are becoming a critical security frontier.

Devcore Team Dominates, Highlighting Complex Chains

The event’s top honors went to the Devcore team, which pocketed a hefty $505,000. Their success wasn’t just about finding one bug; it was about chaining multiple vulnerabilities together to achieve significant impact. This is where the ‘how’ becomes particularly illuminating. Orange Tsai of Devcore, for instance, chained four logic bugs to break out of a sandbox in Microsoft Edge, a feat earning $175,000. Another chain by Tsai, this time three bugs, achieved remote code execution as system on Microsoft Exchange, netting $200,000. This ‘chaining’ is the key. It signifies that attackers aren’t just finding single, obvious flaws; they’re meticulously weaving together less obvious weaknesses to bypass defenses.

This isn’t just about individual software flaws; it’s about the interconnectedness of enterprise systems. A bug in a database might not be critical on its own, but when combined with a flaw in a coding agent that then interacts with that database, the risk profile escalates dramatically. It’s a classic example of emergent complexity in security, where the sum of the parts is far more dangerous than the individual components.

The AI Attack Surface Widens: Why This Matters

So, what does this mean for real people, for developers, for IT security teams? It means the security paradigms we’ve grown accustomed to are already outmoded. We’re no longer just protecting perimeters or individual applications; we’re trying to secure entire AI-powered workflows. The implications for supply chain security are immense. If a coding agent used by countless developers is compromised, the potential for widespread malware injection or data theft is enormous. The 90-day disclosure window ZDI provides to vendors for patching is a tightrope walk – long enough for vendors to act, but short enough to keep threat actors salivating.

The focus on AI databases is also a critical signal. These are the repositories of our data, the training grounds for our models. A breach here could compromise not just intellectual property but also the integrity of AI models themselves, potentially leading to biased or malicious outputs that propagate through an organization. This isn’t a niche problem; it’s becoming foundational.

Is the AI Race Creating New Security Gaps?

There’s a narrative in the tech industry that AI will solve security problems. We’re seeing AI-powered threat detection, AI-driven code analysis. But Pwn2Own Berlin flips that script, showing how AI itself is becoming a potent new attack surface. The race to deploy generative AI and AI-assisted development tools has been so fast-paced that security considerations may have lagged behind. It’s the classic tech industry conundrum: build first, secure later. This event is a loud, clear signal that “later” might be too late.

The fact that researchers are finding multiple, exploitable vulnerabilities in these cutting-edge AI tools suggests a maturity gap. Vendors are pushing out sophisticated AI infrastructure, but the security vetting and hardening process might not be keeping pace with the innovation. This creates a fertile ground for attackers, who are more than happy to exploit the novelty and complexity of these new systems. It’s a Darwinian struggle playing out in the digital realm, with security researchers acting as the natural selection, culling the weak points before malicious actors can.

  • Nguyen Hoang Thach of STARLabs SG exploited VMware ESXi with a memory corruption bug for cross-tenant code execution, earning $200,000.
  • The Devcore team, including Orange Tsai, secured significant winnings by chaining multiple bugs to exploit Microsoft SharePoint ($100,000) and Microsoft Exchange ($200,000).
  • Tsai also achieved a sandbox escape on Microsoft Edge by chaining four logic bugs, earning $175,000.

These aren’t isolated incidents; they represent a systematic uncovering of deep-seated flaws in critical enterprise and AI infrastructure. As vendors scramble to patch these newly revealed vulnerabilities, organizations must also reassess their own security postures, especially concerning the AI tools and services they’ve integrated. The cost of a breach in the AI era could far exceed the prizes awarded at Pwn2Own Berlin.


🧬 Related Insights

Maya Thompson
Written by

Threat intelligence reporter. Tracks CVEs, ransomware groups, and major breach investigations.

Worth sharing?

Get the best Cybersecurity stories of the week in your inbox — no noise, no spam.

Originally reported by InfoSecurity Magazine

Stay in the loop

The week's most important stories from Threat Digest, delivered once a week.