AI 安全是研究人工智能系统面临的安全风险并建立防御机制的研究与工程领域，覆盖对抗攻击、提示词注入、数据中毒、模型窃取、成员推断、后门与偏见歧视等威胁面，核心是把攻防能力纳入模型全生命周期，能力包括红队测试、对抗训练、内容审核与可信部署。它解决的是模型在开放环境中被攻击或产生危害却无防护的问题，并不解决模型自身推理能力不足的问题。
AI 安全的核心是攻防能力的体系化，它可以帮助企业实现以下目标：
让威胁在投产前被主动发现。红队测试以真实攻击手法试探系统，暴露防护盲区而非等问题爆发。评测在固定测试集上按群体分层执行，不同群体的错误率差异被量化记录，差异超阈值即阻断上线。
让训练与推理两侧的入口都被守住。数据侧检测投毒，推理侧拦截注入与越狱，模型资产侧防止窃取。
让合规要求变成可执行的动作。审计日志留存、红队报告与策略阻断点被固化为发布流程的一部分。伦理评审成为发布流程中的一个节点，与安全评审同级，未通过的版本无法进入生产环境。
让安全等级与部署场景匹配。高风险场景适用更严格的审核与人工监督，低风险场景保留必要控制即可。
总之，AI 安全可以帮助企业实现 AI 应用从能跑就能用到可信可审计的转变，把攻防对抗纳入工程流程与治理框架，在开放环境中建立可依赖的安全底线，降低模型被攻击或产生危害带来的业务与合规代价。
翻译
AI safety is the research and engineering field that studies security risks facing AI systems and builds defensive mechanisms, covering threat surfaces such as adversarial attacks, prompt injection, data poisoning, model theft, membership inference, backdoors and bias discrimination. Its core is bringing offensive and defensive capability into the whole model lifecycle, with capabilities covering red-team testing, adversarial training, content moderation and trusted deployment. It solves the problem of models being attacked or causing harm in open environments without protection, but it does not solve the problem of insufficient reasoning capability in the model itself.
At the heart of AI safety is systematizing offensive and defensive capability, which can help companies achieve the following:
Find threats proactively before deployment. Red-team testing probes systems with real attack techniques, exposing defense gaps rather than waiting for incidents.
Guard both the training and inference entry points. Data-side detects poisoning, inference-side blocks injection and jailbreaks, and model-asset-side prevents theft.
Turn compliance requirements into executable actions. Audit log retention, red-team reports and policy checkpoints are fixed as part of the release process.
Match security level to deployment scenario. Higher-risk scenarios get stricter review and human supervision, while lower-risk ones keep necessary controls only.
In short, AI safety can help enterprises shift AI applications from merely running to being trustworthy and auditable, bring offensive-defensive confrontation into engineering processes and governance frameworks, establish a dependable security baseline in open environments, and reduce the business and compliance cost of models being attacked or causing harm.
