The latest ratio of Google model has no key security details, experts say

Thursday, weeks after launching his Powerful AI pattern AI, Gemini 2.5 proGoogle posted a Technical report showing the results of their internal security assessments. However, the be the blind report on the details, expert, make it hard to determine that the pattern may pose.

Technical reports provide useful – and unfiringSometimes – Information that the companies have not ever publicly their ai. For great, acts, see these reports such as good faith efforts to support independent assessments.

Google takes a different combination of some of their revitalians only reporting a pattern “the ignite is also included by the entire” assessions to these written; it resembles those for a separate audit.

Many items experts the techno-still talked by the sparse of the twins of the Gemini 2.5 Report 2,5 Pro, however they will not google Border’s security framework (FSF). I am Google has introduced the FSF last year in what is described as an effort to identify the AI ​​to be able to cause “severe harm.”

“That (report) is very diffresatisfia, contain minimum in the public already available with the public,” Pete Wildeford, Crogeley of the Institute and Strategy, He said to Technique. “It is impossible to check if Corsican live up to their public commitments and so impossible to assess the security and security.”

Thomas Woodside, Co-Found of Texiu sisted that as he is happy a report for Gemini 2.5 Pro, is not convinced of the company’s company company’s Supervisory Company. Woodside has pointed that the last time published the results of dangerous ability was in the June 2022 – for a model announced in February.

Don’t inspire a lot of confidence, Google has not available a report for Gemini 2.5 flashA smaller, most efficient pattern that society announced last week. A portperson said to Techcrough a report for Flash is “soon.”

“I hope this is a promise by Google to start the most frequently”, “Woodside said to include the patternships have committed, since these models had eaten serious risks.”

Google can be one of the first labs ai to propose standardized reports for models, but it is not the only one that has been undedelivering accused about transparency lately. Meta released a SMIMPY security assessment of their new Llama 4 Open Models, and Open Eyed Not to post any report for their Gpt-4.1 series. I am

Hanging up google leader are asseting technology giant made to regulators to maintain a high standard of security of a security of a security and report. Two years ago, Google told the US Government Publising security reports for all models of the “significant” patterns in the purpose. “The society followed that promise with similar commitments at Other countriesproming to “providing public transparency” around the products ai.

Kevin BankSton, a senior board on the ruler to the center for democracy and technology called the tendency of the sporadic and vague the “race to the bottom” in a security.

“Combined with Reports that labs compete have shaved their security test time before the story AI to succeed in the bottom ai

Google told in statements that, while not detailed in their technical reports, carrying out security tests and “the red-security team” to models in front of models.