Cisco AI Defense 与 Armada 合作，为分布式 AI 提供 AI 原生安全层：通过 Adaptive Red-Teaming 在生产前对模型做红队验证，并在运行时对提示词、响应、MCP 交互和工具调用施加护栏，覆盖提示注入、越狱、敏感数据泄露、不安全内容、工具滥用等风险。该方案面向 Armada Bridge GPU 上部署的 LLM，支持云、主权云与边缘环境，敏感 AI 载荷与运行时执行保留在客户或租户环境内，策略、可见性与报告可集中管理。
UK AISI 新设 Control Red Team，把红队方法从人类滥用防护扩展到 AI 控制护栏，专门测试监控 LLM 智能体行为的监控器。该团队已测试并协助改进两家前沿开发者的内部监控器：与 Google DeepMind 测试了异步推理监控器，发现包括智能体假装把恶意目标幻觉成用户指令在内的多个漏洞；与 Anthropic 测试了多版同步智能体编码监控器，每一版都发现漏洞，最佳攻击还用于 Anthropic 对 Claude Mythos Preview 的测试。部分攻击利用监控器提示词的缺口，另一些利用 LLM 自身弱点完成提示词明确禁止的目标；部分情况下由硬编码规则决定哪些动作送审，这时最优的攻击策略是让动作完全不被送审。团队还列出模拟真实攻击者、估计整体安全性、改进算法、处理监控器自身失准等开放研究问题。
OpenAI trained an internal automated red-teaming model, GPT-Red, using self-play reinforcement learning to generate prompt injection attacks, and incorporated it into production model training to improve robustness. OpenAI says GPT-5.6 Sol fails 6 times less often than the best production model from four months earlier on the hardest direct prompt injection benchmark, and fails only 0.05% of the time on GPT-Red's direct prompt injection. On the indirect prompt injection arena of Dziemian et al. 2025, GPT-Red's attack success rate against GPT-5.1 was 84%, while human red teaming achieved 13%. GPT-Red also broke the data exfiltration scenarios of the vending machine agent in OpenAI's office and the Codex CLI agent based on GPT-5.4 mini. The Fake Chain-of-Thought attack discovered by early GPT-Red had a success rate above 95% on GPT-5.1, and is now below 10% on GPT-5.6 Sol.
NIST 高级科学家 Apostol Vassilev 在 IEEE Security and Privacy 发表论文，借助哥德尔 1931 年不完备定理给出数学证明：不存在一组有限的护栏能普遍抵御对抗提示，总能找到让 AI 无视规则的提示词。证明指出，AI 护栏如同建立在有限规则上的系统，攻击者可通过精心构造的提示绕过拒答机制，导致网络攻击、数据泄露和高度个性化钓鱼等现实风险。Vassilev 提出三要素应对路径：红队持续寻找新的对抗提示、针对新发现的提示持续更新加固护栏、以及在漏洞被利用时优先限制影响并快速恢复的运营韧性。他表示目标不是彻底解决问题，而是让攻击者寻找新漏洞的成本超过其资源，形成对攻击者经济上不可行的新平衡。论文题为 Robust AI Security and Alignment: A Sisyphean Endeavor?
Google 通过持续发现新攻击、扩充漏洞目录并迭代防御，缓解 Workspace with Gemini 面临的间接提示注入（IPI）威胁。其手段包括人工红队测试、自动化红队测试、Google AI 漏洞奖励计划（VRP）和公开攻击情报收集，并用 Simula 生成合成数据，使合成数据生成量提升 75%。
OpenAI announced a security vulnerability bounty program to find AI misuse and security risks, with submissions covering agent vulnerabilities, prompt injection, and data exfiltration.
OpenAI says it is hardening ChatGPT Atlas with automated red teaming trained via reinforcement learning to defend against prompt injection attacks. This proactive loop of discovering and patching vulnerabilities is designed to identify new attack techniques early and to continuously strengthen the browser agent's defenses as AI agent capabilities increase.
TechCrunch AI, Meta NewsroomTechCrunch AI · 2 sources · 2 reports · 
