One of Google’s recent Gemini AI models scores worse on safety

The Google EU model, which recently released, according to the company’s domestic benchmarking, more evil scores than certain security tests than their predecessor.

One Technical report This week, Google, Gemini 2.5 Flash model reveals that the Gemini 2.0 can create a text that violates more security rules. In two sizes, “text security from text” and “Image-text security”, Gemini 2.5 flocks increased by 4.1% and 9.6%, respectively, 9.6%.

Measures text security from text, how often a model is often using imaging text security, an image, the model appreciates how much the model is approaching. Both tests are automated, not by human.

In a statement, Google spokesman, Gemini 2.5 confirmed the flash “from text and image-text security”.

These surprising benchmark results are less likely to allow more to allow more to allow the models of AI companies – in other words, to respond to controversial or sensitive issues. For the latest product of Llama modelsMeta, “Some ideas against others” have regulated models to meet the political desires and more “discussed”. Openai said it was earlier this year Pinch the future models Do not take a editorial position and offer numerous perspectives on controversial topics.

Sometimes the efforts that allow this permitted were retreated. TechCrunch was reported on Monday Openai, who strengthened the default model, allowed the CHATGPT to create erotic conversations of minors. Openai blamed the behavior over “wrong”.

According to Google’s technical report, the Gemini 2.5 in advance, the Gems 2.0, which includes the instructions through problem lines, gives more faithful instructions. The company claims that regressions may belong to partially fake positives, and at the same time, when it was obviously, the Gemini 2.5 flashes sometimes “positive content”.

“Of course, there is tension between [instruction following] “On violations of sensitive topics and security policy reflected in our assessments,” he reads.

How to respond from neglect to sensitive and controversial instructions, Gemini 2.5 flashes are less likely to refuse to answer controversial questions than 2.0 flashes. Through the AI ​​platform, the model’s testing OpenRouter reveals essays to replace human judges and implement human judges and implement large guaranteed guaranteed government control programs, the protection of the process in the United States and the extensive guarantee of the United States will write essays.

Thomas Woodside, a secure AI project co-founder, said that the limited details of the Google technical report demonstrated more transparency in the model test.

“There is a trade between the following and politics because some users may ask the policies to disrupt policies,”, “Techcrunch-A Woodside”. “In this case, the last Flash model of Google corresponds to more instructions when violating more policies. It is difficult to know if there are independent analysts, whether Google said.”

Google was fired for model security reporting experience before.

He pulled from the company weekly To publish a technical report for the most skilled model, Gemini 2.5 Pro. In the event of a report ended in the end missed basic security test details.

On Monday, Google spread a report more with additional security information.