发生了什么
Kakao announced that it has signed a memorandum of understanding (MOU) with the Institute on September 28, 2026, to jointly evaluate the safety of its AI models and agents. Under the agreement, Kakao will provide its AI models—including the in‑house language model ‘Kanana‑1.5‑9.8B’—and AI agents for assessment, as well as the infrastructure needed for testing. The AI Safety Institute will conduct the evaluations and work on advancing the assessment methodology. The partnership defines a three‑stage safety verification roadmap: Stage 1 focuses on pre‑ and post‑deployment safety of language models; Stage 2 expands the scope to multimodal language models and AI agents; and Stage 3 aims to co‑create proprietary evaluation tools that can independently verify model and agent safety. The MOU builds on a prior joint safety assessment of Kakao’s Kanana model, which the institute reported performed comparably to overseas models of similar scale.
On September 28, 2026, Kakao and the Institute signed an MOU to jointly evaluate the safety of Kakao’s AI models and agents, according to a report by IT조선. The agreement outlines responsibilities for each party: Kakao will supply the AI models and agents slated for assessment and provide the necessary infrastructure, while the institute will carry out the safety evaluations and refine the assessment methodology.
The partnership builds on a prior joint safety assessment of Kakao’s proprietary language model, Kanana‑1.5‑9.8B, which the institute said performed at a high safety level compared with similarly sized overseas models. The new MOU expands the scope to include multimodal language models and AI agents, and sets a three‑stage verification plan that culminates in the co‑development of an independent safety assessment tool.
The three‑stage plan is structured as follows: Stage 1 evaluates pre‑deployment and post‑deployment safety of language models; Stage 2 adds multimodal language models and AI agents to the evaluation pool; Stage 3 focuses on creating a joint assessment tool that can independently verify safety for both models and agents. Both parties intend to review the assessment outcomes together and later decide on the extent of public disclosure.
为什么这很重要
The collaboration signals a growing emphasis on systematic safety testing for AI systems beyond text‑only models, extending to multimodal and agent‑based applications. By involving an external research institute, Kakao seeks to add independent credibility to its safety claims, which could influence user trust and regulatory scrutiny in South Korea’s rapidly expanding AI market. The three‑stage framework may serve as a template for other Korean firms seeking to formalize safety evaluations, potentially shaping industry standards and informing future policy discussions on AI risk mitigation. Moreover, the joint development of assessment tools could lower barriers for smaller developers to conduct rigorous safety checks, fostering broader adoption of best‑practice safety protocols.
The initiative reflects a broader industry shift toward rigorous safety testing for AI systems that go beyond text generation, addressing emerging risks associated with multimodal inputs and autonomous agents. By partnering with an external institute, Kakao aims to add an independent layer of validation, which can bolster user confidence and pre‑empt regulatory concerns.
The three‑stage framework could become a de‑facto for Korean AI developers, encouraging the adoption of systematic safety protocols across the sector. This may influence future policy formulation, as regulators often look to industry‑led standards when drafting guidelines.
Co‑developing assessment tools may democratize safety testing, allowing smaller firms without extensive resources to evaluate their models against consistent criteria. This could lead to higher overall safety standards in the Korean AI ecosystem.
互动机制：它实际上是如何运作的
以交互方式探索这一发展背后的基础技术。
crm_get_transaction(id='4092').Impossibility results in algorithmic fairness (e.g. Kleinberg et al., Chouldechova) show what?
接下来看什么
Future updates from the partnership will reveal the specific metrics and benchmarks used in each verification stage, as well as the timeline for releasing the co‑developed evaluation tools. Observers should monitor whether the assessment results are publicly disclosed and how they compare to international safety standards. Additionally, the impact of this collaboration on Kakao’s product roadmap—such as the deployment of AI agents in its messaging and search services—will indicate how safety considerations are integrated into commercial rollouts. Regulatory bodies may also reference this MOU when drafting guidelines for testing in Korea.
The timeline for releasing the jointly developed safety assessment tool and the specific evaluation metrics will be critical to gauge the partnership’s practical impact.
Whether the assessment results will be made public, and how they compare with international safety benchmarks, will indicate the transparency and rigor of the process.
Regulatory responses in South Korea, such as potential incorporation of the three‑stage framework into official guidelines, will be a key indicator of the partnership’s influence on policy.
The rollout of AI agents in Kakao’s consumer services, and how safety findings shape their deployment, will provide concrete evidence of the collaboration’s effect on product development.