A bunch of 60 U.Ok. lawmakers has signed an open letter accusing Google DeepMind of violating its commitments to AI security with the discharge of Gemini 2.5 Professional. The letter, printed by political activist group PauseAI, accuses the AI firm of breaking the Frontier AI Security Commitments it signed at a global summit in 2024 by not releasing the AI mannequin with key security info.
At a global summit co-hosted by the U.Ok. and South Korea in February 2024, Google and different signatories promised to “publicly report” their fashions’ capabilities and danger assessments, in addition to disclose whether or not exterior organizations, comparable to authorities AI security institutes, had been concerned in testing.
Nonetheless, when the corporate launched Gemini 2.5 Professional in March 2025, the corporate did not publish a mannequin card, the doc that particulars key details about how fashions are examined and constructed. This was regardless of the corporate’s assertions that the brand new mannequin outperformed rivals on trade benchmarks by “significant margins.” As a substitute, the AI lab launched a simplified six-page mannequin card three weeks after it first made the mannequin publicly obtainable as a “preview” model. On the time, one AI governance professional referred to as this report “meager” and “worrisome.”
The letter referred to as Google’s delay a “failure to honour” the corporate’s dedication on the summit and “a troubling breach of belief with governments and the general public.” The letter additionally took concern with what it referred to as a “minimal ‘mannequin card’” that lacked “any substantive element about exterior evaluations,” in addition to Google’s refusal to substantiate whether or not authorities businesses just like the U.Ok. AI Safety Institute participated in testing.
A spokesperson for Google DeepMind beforehand informed Fortune that any suggestion that the corporate had reneged on its commitments was “inaccurate.” The corporate additionally mentioned in Might {that a} extra detailed “technical report” would come later when it makes a remaining model of the Gemini 2.5 Professional “mannequin household” absolutely obtainable to the general public. The corporate appeared to supply an extended report in late June, months after the total model was launched.
The corporate didn’t instantly reply to Fortune’s request for remark in regards to the latest letter. Nonetheless, as a spokesperson informed Time: “We’re fulfilling our public commitments, together with the Seoul Frontier AI Security Commitments.”
“As a part of our improvement course of, our fashions bear rigorous security checks, together with by UK AISI and different third-party testers—and Gemini 2.5 is not any exception,” they added.
When Google first launched the preview model of Gemini 2.5 Professional, critics mentioned that the lacking system card appeared to violate a number of different pledges the AI firm had made, together with the 2023 White Home Commitments and a voluntary Code of Conduct on Synthetic Intelligence signed in October 2023.
Google wasn’t the one firm to signal these pledges after which seem to drag again on security disclosures. Meta’s mannequin card for its frontier Llama 4 mannequin was about as temporary and restricted intimately because the one Google launched for Gemini 2.5 Professional, and it, too, drew criticism from AI security researchers.
Earlier this yr, OpenAI introduced it might not publish a technical security report for its new GPT-4.1 mannequin. The corporate argued that GPT-4.1 is “not a frontier mannequin,” since its reasoning-focused programs like o3 and o4-mini outperform it on many benchmarks.
The latest letter calls on Google to reaffirm its commitments to AI security, asking the tech firm to outline deployment clearly as the purpose when a mannequin turns into publicly accessible, decide to publishing security analysis stories on a set timeline for all future mannequin releases, and supply full transparency for every launch by naming the federal government businesses and unbiased third events concerned in testing, together with the precise testing timelines.
“If main corporations like Google deal with these commitments as elective, we danger a harmful race to deploy more and more highly effective AI with out correct safeguards,” Lord Browne of Ladyton, a Member of the Home of Lords and one of many letter’s signatories, mentioned in a press release.
