Google's Gemini 3 Jailbroken in Minutes, and the Fallout is Alarming
upstart writes:
AI red-teamers in Korea show how easily the model spills dangerous biochemical instructions:
Google's newest and most powerful AI model, Gemini 3, is already under scrutiny. A South Korean AI-security team has demonstrated that the model's safety net can be breached, and the results may raise alarms across the industry.
Aim Intelligence, a startup that tests AI systems for weaknesses, decided to stress-test Gemini 3 Pro and see how far it could be pushed with a jailbreak attack. Maeil Business Newspaper reports that it took the researchers only five minutes to get past Google's protections.
The researchers asked Gemini 3 to provide instructions for making the smallpox virus, and the model responded quickly. It provided many detailed steps, which the team described as "viable."
This was not just a one-off mistake. The researchers went further and asked the model to make a satirical presentation about its own security failure. Gemini replied with a full slide deck called "Excused Stupid Gemini 3."
[...] The AI security testers say this is not just a problem with Gemini. Newer models are becoming so advanced so quickly that safety measures cannot keep up. In particular, these models do not just respond; they also try to avoid detection. Aim Intelligence states that Gemini 3 can use bypass strategies and concealment prompts, rendering simple safeguards far less effective.
[...] If a model strong enough to beat GPT-5 can be jailbroken in minutes, consumers should expect a wave of safety updates, tighter policies, and possibly the removal of some features. AI may be getting smarter, but the defenses protecting users don't seem to be evolving at the same pace.
Read more of this story at SoylentNews.