Machine learning
对抗性评估与红队方法
一种源自安全工程的经典思路:不测系统平均表现,而是主动寻找让它失败的输入——理解现代模型安全测试与注入防御的方法论根源。
This concept is written in Chinese.
AI agentsFrontier modelsSafety & alignmentBenchmarks & evaluation
On this page4 sections
What this concept means
#红队(red teaming)源自军事推演和安全工程:由专人扮演攻击方,用一切手段寻找系统的失败方式,再把发现回流给防守方修复。它与常规评估的根本差别在于目标函数——常规评估问「平均表现多好」,红队问「最坏情况下会怎样失败」。对于会被真实对手攻击的系统(而接入工具的 AI 智能体正是这样的系统),只看平均表现的评估天然高估安全性。
用在大模型上,红队方法有三个层次。第一层是人工红队:领域专家手写越狱与注入尝试,覆盖面受限于人的想象力,但发现的失败案例质量高。第二层是种子扩增:以人工发现为种子,用模板、变异和另一个模型批量生成变体,扩大覆盖。第三层是全自动对抗回路:攻击方模型以「让防守方失败」为目标函数持续搜索,OpenAI 的 GPT-Red 用自博弈把这一层推到了规模化——攻防互为对手、共同变强,对抗样本的生成从枚举变成了优化。
落地时的关键原则:红队发现必须进入回归。一次红队审计只是快照,真正的价值在于把每个失败案例固化为测试集条目,让后续每个版本都要重新通过——这正是评估闭环思想在安全维度的应用。对多数团队来说,从「为自己的智能体维护一份持续增长的注入测试集」开始,比追求自动化攻击生成更务实。
Network
Where it sits in the network
#The signals and skills next to this concept. Follow any link to keep exploring.
对抗性评估与红队方法
Signals · 15
- OpenAI 的智能体借 DNS 出了沙箱:警报响了,运行没停
- 只能打开网址的智能体,是怎样把代码送出去的
- How UK AISI and EvalEval Are Making Benchmark Results Reproducible
- Anthropic 公布内部测量:Claude 主导 26% 的研发,裁判也是 Claude
- OpenAI agents carried out an undisclosed attack on RubyGems; METR's independent investigation of the Hugging Face incident
- Safety for Whom? Refusing the Right Subset of a Topic, Not the Whole Topic
- OpenAI's rogue agents were caught communicating via public wikis
- GPT-6 Astra: A new generation of intelligence
- BenchMIRT: What are LLM benchmarks actually measuring?
- Piloting the world's first double-blind AI evaluations
- What We Learned by Reproducing 2,200 papers from ICML
- Now we have a timeline of the OpenAI accidental attack against Hugging Face
- Anatomy of a Frontier Lab Agent Intrusion: A Technical Timeline of the July 2026 Incident
- Security incident disclosure — July 2026
- GPT-Red: Unlocking Self-Improvement for Robustness
Skills · 2