Google DeepMind、英国AI Security Instituteとの提携を深化 Deepening our partnership with the UK AI Security Institute

Google DeepMind has announced an expanded partnership with the UK's AI Security Institute (AISI), deepening collaboration on the evaluation and stress-testing of frontier AI models. The move signals continued momentum behind government-led oversight of advanced AI systems in coordination with the labs that build them.

Under the deepened arrangement, AISI will gain richer access to DeepMind's frontier models — including Gemini — to conduct pre-deployment evaluations across risk areas such as cyber offense capabilities, chemical and biological knowledge uplift, and model autonomy. The two organisations also intend to co-develop methodologies for assessing safeguards, sharing technical insights that can feed back into both DeepMind's internal safety processes and AISI's evolving evaluation science.

AISI itself emerged out of the November 2023 Bletchley Park AI Safety Summit, originally founded as the AI Safety Institute before being rebranded earlier this year to emphasise national-security framing. It now sits among a small but growing network of state-backed evaluation bodies, including the US AI Safety Institute housed within NIST, and counterparts forming in countries such as Japan, Singapore and Canada. Cross-border cooperation between these institutes has become a notable feature of the post-Bletchley landscape.

The DeepMind announcement fits a broader pattern: OpenAI and Anthropic have struck similar pre-deployment testing agreements with AISI and its US counterpart, making third-party government evaluation a de facto norm for top-tier frontier systems. In parallel, the EU's AI Act introduces mandatory systemic-risk assessments for general-purpose models above a compute threshold, blurring the line between voluntary commitments and regulatory obligations.

For Google DeepMind specifically, the partnership dovetails with its Frontier Safety Framework, which sets out internal thresholds and mitigations tied to dangerous capability evaluations. Layering external evaluation from AISI on top arguably strengthens the credibility of those self-assessments and provides a degree of public-interest scrutiny that pure internal red-teaming cannot offer. It may also help DeepMind navigate diverging regulatory expectations across the UK, EU and US.

Questions remain, however, about how deeply any government institute can probe models within the tight timelines of commercial release cycles, and how findings will be communicated when evaluators identify concerns. AISI has so far published selective results from earlier rounds of testing, but full transparency is constrained by both security sensitivities and the proprietary nature of the systems involved. Observers are likely to watch closely whether deeper access translates into meaningfully different safety outcomes, or whether it primarily formalises a relationship that was already operating informally.

Either way, the announcement underlines that frontier AI safety is increasingly being shaped not just inside labs, but at the interface between developers and a new class of technical government bodies.