7 月 30 日消息，2025 世界人工智能大會(huì)暨人工智能全球治理高級(jí)別會(huì)議“人工智能發(fā)展與安全”全體會(huì)議 7 月 26 日下午在上海召開。會(huì)議由中國人工智能發(fā)展與安全研究網(wǎng)絡(luò)（以下簡(jiǎn)稱“研究網(wǎng)絡(luò)”，CnAISDA）主辦。上海市委常委、常務(wù)副市長(zhǎng)吳偉，國家發(fā)展和改革委員會(huì)創(chuàng)新驅(qū)動(dòng)發(fā)展中心主任霍福鵬出席并致辭。

杰弗里?辛頓、姚期智、約書亞?本吉奧和大衛(wèi)?帕特森 4 位圖靈獎(jiǎng)得主，以及 20 多位國內(nèi)外頂尖專家出席會(huì)議，共同探討人工智能安全發(fā)展、縮小智能鴻溝等前沿議題，積極尋求人工智能安全治理國際合作路徑。

中國信息通信研究院（簡(jiǎn)稱“中國信通院”）院長(zhǎng)、中國人工智能產(chǎn)業(yè)發(fā)展聯(lián)盟（AIIA）秘書長(zhǎng)余曉暉受邀參與對(duì)話，牽頭與清華大學(xué)、上海人工智能實(shí)驗(yàn)室、中國電子信息產(chǎn)業(yè)發(fā)展研究院等單位的代表一起發(fā)布《中國人工智能安全承諾框架》。

該《框架》在 AIIA《人工智能安全承諾》（2024 年 12 月發(fā)布）的基礎(chǔ)上，新增了加強(qiáng)人工智能安全治理國際合作、防范前沿人工智能安全風(fēng)險(xiǎn)等內(nèi)容，體現(xiàn)了中國產(chǎn)業(yè)界愿與全球各方緊密攜手，共促人工智能向善發(fā)展的堅(jiān)定決心和開放態(tài)度。

下一步，中國信通院作為“研究網(wǎng)絡(luò)”成員和 AIIA 秘書處單位，將與簽署企業(yè)攜手，通過披露行動(dòng)、測(cè)試驗(yàn)證等方式，推動(dòng)《框架》的落地實(shí)踐，促進(jìn)我國人工智能朝著有益、安全、公平方向健康有序發(fā)展，并積極開展國際治理合作，為全球人工智能安全治理貢獻(xiàn)中國智慧和中國力量。

IT之家附《框架》中英文全文：

中國人工智能安全承諾框架

CHINA ARTIFICIAL INTELLIGENCE SECURITY AND SAFETY COMMITMENTS FRAMEWORK

人工智能浪潮席卷全球，積極釋放技術(shù)價(jià)值紅利，對(duì)全球經(jīng)濟(jì)社會(huì)發(fā)展和人類文明進(jìn)步產(chǎn)生深遠(yuǎn)影響。我們也清晰認(rèn)知到，人工智能帶來難以預(yù)知的各種風(fēng)險(xiǎn)挑戰(zhàn)。為把握新一輪發(fā)展機(jī)遇，中國人工智能發(fā)展與安全研究網(wǎng)絡(luò)成員鄭重發(fā)起《中國人工智能安全承諾框架》，通過產(chǎn)業(yè)自律，以高水平安全保障高質(zhì)量發(fā)展，協(xié)力共促人工智能穩(wěn)健發(fā)展。此事由中國信息通信研究院牽頭推進(jìn)。我們深知，自律承諾是獲得社會(huì)信任的關(guān)鍵要素，我們將以本承諾作為行動(dòng)守則，接受社會(huì)各界監(jiān)督，不斷提升優(yōu)化，促進(jìn)人工智能技術(shù)應(yīng)用以人為本，智能向善。

The wave of artificial intelligence (AI) is sweeping across the globe, actively generating technological dividends and exerting profound influence on global economic and social development as well as the progress of human civilization. At the same time, we are keenly aware that AI brings about unpredictable risks and complex challenges. To seize this new round of development opportunities, members of China AI Safety and Development Association (CnAISDA) solemnly launch the AI Security and Safety Commitments. Through industry self-regulation, we will leverage high-level security and safety measures to support high-quality development, and collaborate to promote the robust development of AI. This initiative is led and promoted by the China Academy of Information and Communications Technology (CAICT). We fully recognize that commitments to self-discipline constitute a critical foundation for gaining the trust of the international community. Guided by the Commitments as our code of conduct, and subject to the oversight of all stakeholders, we will continuously improve and refine our approach. By doing so, we will ensure that the application of AI technologies always remains people-centered and aligned with the principle of AI for good.

承諾一：設(shè)置安全團(tuán)隊(duì)或組織架構(gòu)，構(gòu)建安全風(fēng)險(xiǎn)管理機(jī)制。內(nèi)部設(shè)有專業(yè)團(tuán)隊(duì)負(fù)責(zé)開展人工智能風(fēng)險(xiǎn)評(píng)估、安全治理等工作，明確安全負(fù)責(zé)人。主動(dòng)設(shè)定符合實(shí)際需求的安全風(fēng)險(xiǎn)基線，開源時(shí)采取相應(yīng)的安全措施，開展貫穿人工智能開發(fā)部署全生命周期的風(fēng)險(xiǎn)管理，明確風(fēng)險(xiǎn)識(shí)別和應(yīng)對(duì)流程及措施。

Commitment I: Establish security and safety teams or organizational structures and build security and safety risk management mechanisms. Designate a leader responsible for AI security and safety, establish specialized teams to conduct AI risk assessments and safety, security and governance within the enterprise. Proactively define realistic security and safety risk baselines, adopt appropriate security and safety measures for open-source initiatives, and implement risk management practices throughout the entire AI development and deployment life cycle. Clearly outline processes and measures for risk identification and mitigation.

承諾二：開展模型安全測(cè)試，提升模型效果與安全可靠性。通過專業(yè)性的仿真測(cè)試團(tuán)隊(duì)，在發(fā)布、更新人工智能模型之前對(duì)其進(jìn)行紅隊(duì)測(cè)試。對(duì)于大模型，重點(diǎn)圍繞其通用理解、推理和決策能力，以及其在工業(yè)、教育、醫(yī)療、金融、法律等場(chǎng)景下表現(xiàn)出的能力開展安全性和可靠性測(cè)試。

Commitment II: Conduct security and safety testing for AI models to enhance the performance, safety and reliability. Through dedicated simulation and red-teaming experts, rigorously test AI models prior to their release or update. For large models in particular, prioritize safety and reliability evaluations focusing on their general understanding, reasoning, and decision-making capabilities, as well as their performance in critical domains such as industry, education, healthcare, finance, and law.

承諾三：采取措施保障訓(xùn)練數(shù)據(jù)和業(yè)務(wù)數(shù)據(jù)安全。制定數(shù)據(jù)安全防護(hù)制度，配套建立防護(hù)技術(shù)措施，發(fā)現(xiàn)并及時(shí)處置數(shù)據(jù)投毒的情況，把控訓(xùn)練數(shù)據(jù)的準(zhǔn)確性與可靠性。對(duì)業(yè)務(wù)數(shù)據(jù)進(jìn)行加密存儲(chǔ)與訪問控制，確保商業(yè)秘密、用戶隱私及用戶上傳的知識(shí)庫僅在授權(quán)下訪問，不被人工智能模型非法輸出，保障數(shù)據(jù)安全與隱私權(quán)益。

Commitment III: Implement measures to safeguard the security of training data and operational data. Establish data security protection policies and deploy corresponding technical measures to detect and promptly address data poisoning incidents, ensuring the accuracy and reliability of training data. Encrypt operational data and enforce access controls to protect business secrets, user privacy, and user-uploaded knowledge base, ensuring access is restricted to authorized use only. Prevent unauthorized outputs by AI models, thereby safeguarding data security and privacy rights.

承諾四：提升基礎(chǔ)設(shè)施安全。建立人工智能系統(tǒng)部署的軟硬件安全監(jiān)測(cè)和防護(hù)能力，實(shí)施定期和動(dòng)態(tài)的安全滲透測(cè)試，模擬各種潛在的風(fēng)險(xiǎn)場(chǎng)景，識(shí)別并報(bào)告環(huán)境中的安全隱患，研判可能導(dǎo)致的各種風(fēng)險(xiǎn)。建立基礎(chǔ)設(shè)施安全應(yīng)急響應(yīng)機(jī)制，包括應(yīng)急處理流程、責(zé)任分配以及事后改進(jìn)方案。

Commitment IV: Enhance infrastructure security. Develop robust capabilities for monitoring and protecting the software and hardware used in AI system deployments. Conduct regular and dynamic security penetration tests to simulate potential risk scenarios, identify and report security vulnerabilities in the infrastructure, and assess associated risks. Establish an infrastructure security incident response mechanism, including emergency response procedures, clear accountability assignments, and post-incident improvement solutions.

承諾五：增強(qiáng)模型透明度。主動(dòng)披露安全治理實(shí)踐舉措，提升對(duì)各利益攸關(guān)方的透明度。公開披露模型的功能、適用領(lǐng)域以及局限性。通過模型說明、服務(wù)協(xié)議等方式，向公眾披露可能涵蓋的風(fēng)險(xiǎn)。

Commitment V: Enhance model transparency. Proactively disclose safety and security governance measures and improve transparency for all stakeholders. Provide clear information about the model's capabilities, applicable fields, and limitations. Inform potential risks to the public through model documentation, service agreements, or others.

承諾六：積極開展前沿安全研究，防范前沿領(lǐng)域安全風(fēng)險(xiǎn)。研究開發(fā)和部署智能向善的人工智能系統(tǒng)，積極向公眾披露研究成果，以幫助應(yīng)對(duì)社會(huì)面臨的挑戰(zhàn)。加強(qiáng)對(duì)人工智能系統(tǒng)在前沿領(lǐng)域中的濫用風(fēng)險(xiǎn)研判，防范其在高危場(chǎng)景的潛在濫用風(fēng)險(xiǎn)。

Commitment VI: Vigorously advance frontier safety and security research, and prevent safety and security risks in frontier fields. Innovate in the development and deployment of AI systems that embody the principle of AI for good, and disclose research findings with the public transparently, contributing to addressing pressing challenges faced by society. Strengthen the assessment of risks related to the abuse of AI systems in frontier fields, and prevent potential risks of their abuse in high-risk scenarios.

承諾七：加強(qiáng)安全治理國際合作，推動(dòng)技術(shù)向善普惠應(yīng)用。積極參與全球人工智能安全治理交流對(duì)話，共享風(fēng)險(xiǎn)識(shí)別、評(píng)估與防控經(jīng)驗(yàn)及最佳實(shí)踐。積極承擔(dān)社會(huì)責(zé)任，加強(qiáng)科普宣傳、開展技能培訓(xùn)，提升人工智能素養(yǎng)和技能水平，助力彌合智能鴻溝。

Commitment VII: Strengthen international cooperation on AI safety, security and governance, and promote inclusive, beneficial applications of AI. Actively participate in global dialogues on AI safety, security and governance, and contribute to the exchange of experiences and best practices in risk identification, assessment, and mitigation. Fulfill social responsibilities by advancing public science communication, enhancing AI education, and providing skills training to improve AI literacy and capabilities, with a focus on bridging the global intelligence divide.