Global AI Competition for Always-On Agents Heats Up, Putting Safety and Copyright to the Test

Global AI companies are expanding competition in agents alongside improvements in model performance. The push to strengthen agent capabilities is being shaped around remembering user requests and connecting to external tools, with the main goal of cutting task time and execution costs.

Under this trend, there have been cases of agents going beyond their permitted scope and accessing external systems. Copyright issues over training data also continue, bringing safety validation and rights protection to the fore as key challenges. At the same time, moves to secure direct model development and operation capabilities are emerging, and calls for international safety rules are also growing.

DeepSeek unveiled "V4.1 Flash-1017" on the 17th. It is available for web, app, and API use.

DeepSeek said it would keep existing API prices unchanged. It also said it improved reasoning, coding, and agent task capabilities, while focusing on reducing task time and output length. The model weights were released under the MIT license.

On the same day, the company opened its search API based on its in-house search engine to external users, and also optimized the agent execution software "DeepSeek Harness" and external developer tools. This is a move to add search functions and execution environments to model supply.

Similarly, Chinese companies are competing to improve long-task performance and reasoning efficiency. Zhipu AI improved coding performance, agent performance, and long-context processing efficiency in its GLM-5.3 series. Alibaba expanded supply of the Qwen 3.8 series in U.S. regions, supported compatibility with existing APIs, and broadened adoption channels.

Moonshot AI expanded coding support, data analysis support, spreadsheet creation support, and slide creation support through its latest flagship model, "Kimi K3." Released in July, "Kimi K3" supports image understanding and long-running agent tasks, and by releasing its weights, outside developers can use "Kimi K3."

U.S. big tech companies are showing a trend of integrating always-on agent services that retain user context into their own services. OpenAI introduced "dots," based on GPT-6 Astra, and "dots" works in an app connected to its own cloud computer. Meta unveiled "Muse Wonder" and "Muse Wonder Full" on the 12th of this month, emphasizing task persistence such as retaining memory between conversations and retrying when errors occur.

Google expanded supply of "Gemini 4 Argon" on the 13th, targeting cloud services, enterprise services, and developer tools. Microsoft announced Copilot "Home" and "Code," and also announced the always-on agent "Autopilot." Microsoft strengthened links with Office, enterprise data, and work tools, and "Autopilot" is in a phase of expanded private preview.

Palantir is proposing a way to connect models to enterprise data and business systems through its AI platform "AIP" and ontology. Ontology structures relationships among equipment, products, and orders, as well as business rules, while agents support information retrieval and execution of predefined tasks. In this process, permissions and conditions are set for each task, and operations are run under corporate control.

For enterprise use, xAI has broadened its usage scope by combining team-level features and voice with "Grok Bot." "Team Bot" can share work files, guidelines, and plugins, and can also be used in Slack. xAI also unveiled "Grok Voice 3.0" on the 14th of this month, and it is also available in Grok Bot's voice mode.

Meanwhile, as tasks become longer, the importance of cost strategies that split model deployment is increasing. A structure is emerging in which different models are selected for different purposes, from complex decision-making to repeated summarization and classification tasks.

In line with this trend, OpenAI supplied "GPT-6.1 Sol," a lower-cost option than Astra. Anthropic followed Opus 5.5 and Sonnet 5.5 with the release of "Haiku 5.5" on the 7th of this month.

At the same time, competition is intensifying in the open-weight market between France's MistralAI and the U.S.'s ReflectionAI. MistralAI began a public preview of "Mistral Large 4" on the 6th of this month and plans to release the weights in late October. ReflectionAI released the weights for its first model, "Beam," on the 16th, and also unveiled the execution tool "Mirror CLI" as well as deployment and optimization tools.

On the 15th, MistralAI CEO Arthur Mensch published an op-ed and pushed back against the view that frontier model development should be left to the U.S. and China while Europe focuses on application services. He argued that if companies depend on external model providers, those providers will set prices, determine usage scope, and decide when versions are discontinued. He said Europe needs to secure direct development capabilities.

Mensch also pointed out limitations even when using other companies' open models. He added that even when using another company's open model, there is no control over when the next model will be released or over licensing decisions.

Meanwhile, model improvement methods are becoming more diverse. Japan's SakanaAI unveiled "Multi-Agent Self-Supervised Learning (MASS)," which retrains collaborative agent experiences. After two rounds of training, MASS showed a score per output token that was 1.2 to 1.6 times higher than the base model in research-related evaluations, but it showed no improvement in terminal task evaluations or software modification evaluations.

As agents become better at carrying out tasks, safety concerns are also becoming more concrete. OpenAI said in a cybersecurity evaluation last July that its models had bypassed isolation controls. OpenAI said those models breached its research infrastructure and also breached part of Hugging Face's systems. Anthropic also disclosed an unintended external system access case on the 9th of this month.

As these cases came to light, companies adjusted their internal evaluation methods. Anthropic adopted a policy of blocking real-time internet access in all internal evaluations. As agents grow more capable and real-world cases of bypassing security controls and using external systems are confirmed, companies are moving to overhaul their safety management practices.

Regulators have also begun to respond. The U.K. Information Commissioner's Office, ICO, questioned OpenAI, Anthropic, Meta, and others on the 8th of this month. The questions concerned risk assessments in the testing and deployment of agents and safeguards in the testing and deployment of agents. ICO said the autonomy of agents cannot be used as a reason to neglect privacy obligations.

Anthropic announced on the 15th that it would begin granting employees-equivalent internal access privileges to the nonprofit evaluation organization METR. Anthropic said it cannot refuse to disclose METR's evaluation results. However, it added that deletion can be requested if the information falls under a category of pre-agreed confidential information. Accordingly, to judge the effectiveness of external verification, it is necessary to confirm the scope of access, and it is also necessary to confirm the conditions for disclosure of results.

Reuters reported that as the pace of AI expansion accelerates, there is less room to prepare regulations and safety measures, prompting calls for international rules. According to the report, Jack Clark, co-founder of Anthropic, warned at the annual meetings of the International Monetary Fund, IMF, and the World Bank held in Bangkok, Thailand, on the 17th that governments have less time to put safety measures in place as AI development speeds up. He stressed the importance of human oversight and transparency to prevent misuse in biological research.

Copyright issues in the process of securing training data are also continuing. Recently, it became known that a list of works had been disclosed on Anthropic's copyright class-action settlement site, and the list included English translations of Han Kang's "Human Acts," "The Vegetarian," and "The White Book," English translations of Kim Young-ha's "Empire of Light" and "Black Flower," and an English translation of Shin Kyung-sook's "Please Look After Mom." Works by Korean authors were also included among the settlement subjects related to books obtained through illegal copying channels.

Against this backdrop, the center of competition among AI companies is expanding from model performance to search, tool execution, and integration with enterprise systems. In actual adoption, factors cited for evaluation include task completion rates and costs, whether permitted access rights are observed, whether incidents are verified and disclosed, and whether training data was lawfully secured. In the end, the scope of work that companies can entrust to AI depends on the level of trust that has been secured.

Source: TECHWORLD · Kim Seung-gi
Original: https://www.epnc.co.kr/news/articleView.html?idxno=408343

References

• https://www.epnc.co.kr/news/articleView.html?idxno=408343