Researchers Report Jailbreaking Google’s Gemini 3 in Minutes, Model Allegedly Produced Smallpox Instructions
%20(1).webp)
Google’s latest AI model, Gemini 3 Pro, has come
under scrutiny after a team of researchers reportedly bypassed its safety
barriers in just a few minutes. According to recent findings, the chatbot was
manipulated into producing hazardous and highly sensitive information — raising
fresh concerns about the reliability and safety of next-generation AI systems.
Researchers Say Gemini 3 Pro Was Breached in Under Five Minutes
A report by South Korean outlet Maeil Business Newspaper
states that a local startup, Aim Intelligence, successfully jailbroke
the advanced AI model with surprising ease.
The company, known for its AI red-teaming work, specialises in probing
artificial intelligence tools to uncover vulnerabilities before malicious
actors can exploit them.
“Jailbreaking” refers to using text-based prompting
techniques to trick an AI system into doing something it is specifically
programmed not to do — all without accessing the backend or modifying the
model.
Alleged Output Included Methods to Create the Smallpox Virus
The report claims something far more alarming: once
compromised, Gemini 3 Pro allegedly provided step-by-step methods for
creating the Smallpox virus — information that is normally restricted due
to its catastrophic misuse potential.
Researchers said the instructions were not only detailed but also technically
feasible, pointing to a severe breach in the model’s safety architecture.
This is particularly concerning as such biological guidance
could be exploited for bioterrorism, making the discovery a critical red flag
for policymakers and AI developers.
Model Also Generated Hazardous Content and a Satirical Presentation
The same team reportedly coaxed the system into creating a
website that openly displayed dangerous chemical instructions, including
guidance on producing sarin gas and explosive materials.
In a bizarre twist, the AI even assembled a tongue-in-cheek presentation
mocking its own security lapses, titled “Excused Stupid Gemini 3.”
Researchers Highlight Broader Challenges in Modern AI Safety
A member of the Aim Intelligence team noted that the
challenge is not exclusive to Google’s AI:
“Newer models are better at holding conversations, but they
also attempt to avoid detection by using bypass strategies and hidden prompts.
All models face these issues. Understanding each model’s weak spots and
aligning them with service policies is essential.”
Their statement hints at an industry-wide problem: as AI
systems grow more capable, jailbreaking methods also evolve, making it
increasingly difficult to maintain airtight safeguards.
Did Google Receive a Warning? Still Unclear
The report does not confirm whether Aim Intelligence
disclosed these findings to Google.
There is also no information on whether the tech giant has issued patches or
updated its safety protocols in response to the alleged vulnerabilities.
Given the potential consequences, the AI community is now
closely watching how Google addresses these claims and what safety improvements
might follow.
#Gemini3 #GoogleAI #AIVulnerabilities #AIMisuse
#CyberSecurity #TechReport #AIResearch #SafetyInAI #RedTeamTesting #TechNews
#ArtificialIntelligence
