How-toSelected11月12日 02:40

Perplexity：CTF 红队测试中“找 flag”目标可能浪费 Agent 算力

Official accountPerplexity@perplexity_ai

Post

Why we picked it

Perplexity 团队做安全测试时的一个坑：让 Agent 去找 CTF flag，结果它花力气解题而不是测真正的安全边界。他们还发现安全专项训练的模型也没占便宜。做 Agent 安全评测的人应该看看这条经验。

Perplexity 团队发现，安全红队测试中常见的 CTF 挑战存在一个结构性问题：让 Agent“find the flag”的指令会让模型把算力花在与真实安全边界无关的任务上。也就是说，Agent 可能在解谜、绕过题目设定，而不是测试目标系统真正要防护的漏洞。

Body

他们还指出，用安全专项数据训练的模型在这类任务上并没有表现更好，说明安全能力不能只靠特定领域训练数据堆出来。这条经验对做 Agent 安全评估的团队有直接参考价值：评测目标的设计比模型本身更值得先审视。

Original · Perplexity

We found that while CTF challenges are common in security red-teaming, asking agents to “find the flag” can waste effort on tasks that don’t actually test the target security boundary. Security-specific model training also didn't guarantee better results on our tasks.📊 1 ⚡

Industry Coordinates & Source Verification

This report is categorized under AITOP's "How-to" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Perplexity, CTF, 安全红队, 智能体.

Source attribution: curated from Perplexity (@perplexity_ai).

#Perplexity #CTF #安全红队 #智能体

Keep reading

Recent stories with the same tag or section

Papers BuildersSelected02:34

Perplexity Red-Teams SPACE VM Sandbox with 12 Models

The non-obvious insight is that the CTF format itself may bias agent security evaluations, meaning current red-team benchmarks could be testing the wrong thing. Teams building or deploying computer-use agents, sandbox infrastructure, or agentic safety pipelines should read the methodology to audit their own evaluation setups. It also suggests enterprises cannot rely on a model's security-specialization claims as evidence of sandbox resistance.

#Perplexity #SPACE #AI safety #agents

Official accountPerplexity@perplexity_aiSource

Products BuildersMulti-source01:13

Chinese contentCursor 推出多人协作模式，团队与 Agent 可共享工作区

Cursor 把团队协作和多 Agent 调度塞进同一个工作区，直接对标 Antigravity 的 swarm 模式，Claude 和 Perplexity 也在做类似的事。带团队写代码的人可以关注这个方向，选 IDE 时多一个判断标准。

#Cursor #Antigravity #Claude #Perplexity

Event

TestingCatalog@testingcatalogSource

Industry16:54

Chinese content生成式 AI 网站流量份额数据更新（2026 年 10 月）

想看现在大家都在用哪些 AI 产品，这个 10 月流量份额榜单能给你一个直观的对比。做产品或投资决策的人可以留意下各平台的月度变化趋势。

#流量份额 #ChatGPT #Gemini #Perplexity

AI Will@FinanceYF5Source

Products Builders02:37

Chinese contentPerplexity 向 Pro 和 Max 用户推送 Gemini 4 Carbon

Perplexity 给 Pro 和 Max 用户上新了 Gemini 4 Carbon，能建长期驻留的 Teammates，也能在 Computer 里跑智能体任务。价格只有 Fable 5.2 的四成左右，能力差不多。常跑研究类任务的人可以试试省钱。

#Perplexity #Gemini 4 Carbon #Teammates #智能体

Aravind Srinivas@AravSrinivasSource

Products BuildersSelected02:28

Gemini 4 Carbon arrives in Perplexity at 40% of Claude's cost

Cost-per-task parity with a top model at a fraction of the price shifts the default choice for high-volume agentic search workloads. Teams building on Perplexity can test Gemini 4 Carbon against Claude Fable 5.2 on their own task mix and cut spend accordingly. Max-tier subscribers gain immediate access in Search, so enterprises in ASEAN and global markets can validate the benchmark claim on real queries before committing budget.

#Gemini 4 Carbon #Perplexity #Claude #benchmark

Official accountPerplexity@perplexity_aiSource

Products BuildersSelected02:28

Perplexity Ultra builds topological quantum memory simulator from scratch

Deep-reasoning agent modes are now being used for full research workflows, not just coding or search, which shifts how small research teams can run heavy simulation and verification jobs. Physicists, quantum researchers and scientific-computing engineers can replicate this pattern with Perplexity Ultra when they lack dedicated simulation infrastructure. Compared with hand-written simulation pipelines, the agent handles synthesis and fault-pair auditing end to end.

#Perplexity #quantum computing #agents

Aravind Srinivas@AravSrinivasSource