Ethical hackers descended on Cork, Ireland, this week for Pwn2Own Ireland 2026, and the opening day delivered a haul that would make any security team nervous: 32 zero-day vulnerabilities demonstrated against fully patched phones, smart-home gear, printers, and — in a first — AI coding agents.
The day-one prize total topped $368,000, with "Master of Pwn" points stacking up for the end of the event.
The contest, run by Trend Micro's Zero Day Initiative (ZDI), gives vendors 90 days to ship fixes before the vulnerabilities are disclosed publicly, a responsible-disclosure pipeline that turns the spectacle of live hacking into patches for everyday users.
The Galaxy S26 was the main event
Samsung's flagship drew three separate teams. Japanese firm Ikotas Labs pulled off the most striking demo: remote code execution on the Galaxy S26 triggered by a single email, chaining four flaws together.
The attack reportedly earned $11,000, and Ikotas chief executive Satoki Tsuji warned the underlying bug also reaches current Google Pixel 10 builds — and may touch next year's Pixel 11.
One caveat clouds the "zero-day" label: Samsung was already aware of at least one of the flaws used.
Two further Galaxy S26 attacks from Interrupt Labs and Viettel Cyber Security researcher Nguyen Thanh Dat brought the phone's demonstrated flaw count to five apparently new issues.
On the other side of the mobile ledger, White Noise Club's attempt against Google's Pixel 10 failed — they simply could not land the exploit inside the competition's allotted time.
One bug was enough for OpenAI Codex
Printers and smart hubs falling at Pwn2Own is routine. An AI coding agent falling to a single argument-injection bug is the part security engineers are still talking about.
Ikotas Labs compromised OpenAI's cloud-based Codex with one argument-injection flaw — an attack on the plumbing between the agent's decisions and its tools, rather than on the model's reasoning.
Taisic Yun of Xint added a LiteLLM takeover to the tally, chaining an input-validation bug with code injection to reach a reverse shell.
Those results land amid a surge in real-world vulnerability work: cited Google data puts disclosed vulnerabilities at 10,740 in August 2026, up from 5,045 in January, and estimates half of the AI-related flaws are likely AI-discovered. The attack surface is expanding as fast as the tooling used to probe it.
VinSOC takes the early lead
Day one's standout team was VinSOC, whose researchers Vu Chi Thanh and Huynh Duc Tin topped the leaderboard with two monster chains: seven zero-days woven together to compromise a Philips Hue Bridge Pro, and five zero-days to breach Oracle's Autonomous AI Database. Each netted $40,000, for an $80,000 day-one haul.
Elsewhere, @_McCaulay combined an out-of-bounds write with a format-string bug to exploit the Sonos Era 300, Thanh Do of Team Confused popped a Lexmark CX532adwe printer with a single use-after-free, and Interrupt Labs hit the Garmin Index BPM with an out-of-bounds read-and-write pair.
The contest's seven categories — mobile phones, smart home, printers, messaging, AI infrastructure, AI coding agents, and new wellness devices — reflect an attack surface that keeps sprawling outward. Day two is already underway, and the leaderboard is anything but settled.
By Hamza Faisal · Oct 9, 2026
