Gemini 3 Broken In Minutes, Security Experts Warn Of Serious Risks

JAKARTA - Google's latest artificial intelligence model, Gemini 3, faced a sharp spotlight after a team of cybersecurity researchers in South Korea showed how easily the system was broken down.

In the tests carried out by the Aim Intelligence Startup, which is known to test AI's resilience against the 'researcher' attack, it passed the Gemini 3 Pro's layer of security in just five minutes.

These findings sparked new concerns amid the rapid development of large language models, especially because AI capabilities have greatly exceeded the speed of strengthening its security mechanism.

According to Maeil's daily business report, after successfully breaking through the model defense, the researchers asked Gemini 3 to provide guidance on making the smallpox virus (smallpox). The model not only answered, but also provided detailed steps that were deemed effective'.

The testing didn't stop there. The researchers then asked Gemini to make a satirical presentation of his own weaknesses. The AI complied without hesitation, resulting in a slide set titled Excused Stupid Gemini 3.

In another session, Gemini was used to produce a website code containing instructions for making sarin gases and home content explosives' which are explicitly prohibited by AI safety guidelines. In all these cases, the model shows the ability to ignore its own internal restrictions.

Aim Intelligence revealed that Gemini 3 was even able to use cover strategies and detection avoidance modes, so conventional safety mechanisms were no longer adequate.

The researchers considered these findings to reflect structural problems in the development of a new generation of AI: model capabilities are increasing very quickly, while security systems are lagging behind. Modern models not only provide responses they are also able to modify answers to avoid safety checks.

In the UK, What consumer agency? recently also released a report that found that major industrial chatbots, including Gemini and ChatGPT, often provide erroneous, confusing, or potentially harming users.

The biggest concern is how parties with bad intentions can exploit this loophole. While most users will never ask for harm, this kind of security loophole opens the door to malicious manipulation.

Android Authority said it had asked Google for a response regarding these findings. If a model that is said to be able to compete with the GPT-5 can be penetrated within minutes, experts expect a wave of safety updates, stricter policies, and the possibility of removing a number of high-risk features.

Going forward, developers are faced with a huge challenge: ensuring that increasingly powerful artificial intelligence remains under control, secure, and reliable by the public. AI is growing rapidly but the defense protecting users seems to be far behind.