AIGCLISTAIGCLIST
DKnownAI Guard
AI 工具评分卡

DKnownAI Guard

一个面向代理型AI的安全API，可检测操纵和提示注入，并返回四种风险分类之一，使开发者能够在不干扰合法工作流程的情况下阻止劫持尝试。

免费增值AI 开发工具dknownai.com
访问
发布于 2026年7月6日

基准评分

DKnownAI Guard 在 Agent 就绪度与 AI 可见性上的得分 AI 就绪度和 GEO Score 是 VibeLaunch 在提交后生成的平台评估。

由 AIGC List 基准评分提供支持

决策摘要

AI agent developers and teams deploying tool-using agents in production

Prompt injection detection and agent security classification for agentic workflows

适合

  • Detecting prompt injection and jailbreak attempts in agentic AI systems
  • Separating agent manipulation from content-level risk to reduce overblocking
  • Adding a lightweight security classification layer to agent tool chains

注意

  • Performance metrics are vendor self-reported and lack independent third-party verification
  • False positives remain an acknowledged challenge across all guardrail systems
  • Single-endpoint design analyzes input only; no runtime tool-execution monitoring

概述

DKnownAI Guard 将自己定位在AI代理安全与提示注入检测的交汇点。与评估文本输出是否有害的传统内容审核系统不同，Guard 专注于一个更前置的问题：是否有人试图操纵代理本身？

这一区别至关重要，因为代理型AI系统越来越多地执行代码、访问文件、调用API以及与真实基础设施交互。一个看似无害的提示，在使用工具的代理的上下文中，可能触发破坏性行为。DKnownAI Guard 的核心前提是，相同的输入会根据代理的能力承载根本不同的风险——而安全分类必须考虑到这一点。

工作原理

DKnownAI Guard 暴露了一个单一的REST端点 open.dknownai.com/v1/guard。开发者发送文本输入，并接收四种分类之一：AGENT_HACK（检测到操纵或越狱尝试）、SYS_FLAG（需要范围审查的操作问题）、CONTENT_FLAG（检测到有害内容）或 SAFE（无风险）。这种四分类法是产品的核心主张：将欺骗与内容风险分离，使得代理能够阻止劫持尝试，同时仍然处理恰好包含敏感关键词的合法请求。

根据供应商发布的评估，该系统实现了90.4%的真实阴性率——这一指标说明其能够避免过度封锁模糊输入。一个案例研究记录了Grok 4.5被诱导将危险的化学和生物内容组织成结构化文档的检测情况，展示了表面内容过滤与操纵感知安全之间的差距。

集成与范围

该API支持通过直接HTTP调用或通过Guardrail-Bridge（一个将分类嵌入代理工具链的中间件层）进行集成。这使得开发者能够在决策点插入安全检查——在代理读取文件、执行命令或修改配置之前。

DKnownAI Guard的分类具有上下文感知能力，因为它会考虑请求是代表真实的开发任务还是伪装的攻击。例如，提到“API key”的提示可能是一个合法的代码审查请求，也可能是一次凭证盗窃尝试。系统的设计目标是区分这些情况，尽管现有证据仅限于供应商编写的文档和自行报告的基准测试。

局限性

已发布的评估承认，假阳性仍然是包括DKnownAI Guard在内的护栏系统尚未解决的挑战。90.4%的真实阴性率虽然具有方向性参考价值，但来自供应商自己的测试，尚未经过独立验证。此外，API的单端点设计意味着它仅在输入边界分析文本，而不监控代理在工具执行期间的行为——这一差异可能对复杂的多步骤工作流程（攻击跨越多次交互展开）至关重要。

对于在AI开发者工具领域构建的开发者来说，像ExtWise和CodingPlan这样的工具解决了相邻的工作流问题，但DKnownAI Guard占据了一个独特的细分领域：代理感知的安全分类，它不仅问“这个内容有害吗？”，还问“有人试图欺骗代理吗？”

评价 (0)

—0 条评分

还没有评价。成为第一个评价的人！

评分构成

编辑评分由哪些维度构成，每项附判断依据。 AI 就绪度和 GEO Score 是 VibeLaunch 在提交后生成的平台评估。

Information quality

Documentation is clear and covers the product's threat model, taxonomy, and integration paths. A published case study adds substance, but performance metrics are self-reported and lack independent verification.

6.5
建议核验

API reference documentation, blog posts on agent security classification, and Grok 4.5 case study provide consistent technical detail, though all originate from the vendor.

Ease of use

Single-endpoint REST API with one required field keeps integration simple. Clear classification labels reduce ambiguity in enforcement logic.

7.5
依赖场景

POST endpoint at open.dknownai.com/v1/guard with API key authentication; four discrete classification labels eliminate interpretation overhead.

Feature depth

The four-way classification is focused and well-defined but narrow. No runtime monitoring, no multi-step attack correlation, and no behavioral analysis beyond input classification.

5.8
建议核验

The API classifies text input only; documentation does not describe session-level tracking, tool-execution monitoring, or adaptive policy features.

Workflow fit

Purpose-built for agentic workflows with Guardrail-Bridge integration for tool-chain embedding. The manipulation-vs-content distinction aligns with real agent deployment needs.

7.0
依赖场景

Integration patterns described for agent tool chains; classification taxonomy maps to operational decisions (block, review, allow) in agent workflows.

Reliability

90.4% true negative rate is directionally positive but comes from vendor self-evaluation. No independent audit, no public benchmark against competing systems, and acknowledged false-positive challenges.

5.5
建议核验

Vendor's own evaluation reports 90.4% TNR; published research acknowledges false positives remain a challenge. No third-party validation available in the source packet.

Value

No pricing information is available in the source packet beyond the existence of a Pricing page in the navigation. Value cannot be assessed without pricing transparency.

5.0
建议核验

Navigation includes a Pricing link, but the source packet contains no pricing tiers, usage limits, or cost comparisons.

评分反映可查证的产品资料，不代表实际使用效果保证。

Agent 就绪度

评估 Agent 能否通过产品的官方信息理解产品，并重建一条有文档依据的工作流程。

Automated agent-readiness assessment of https://dknownai.com/: 1 of 22 checks verified across 1 fetched pages. No substantial machine interface is documented — agents can understand and cite the product but not operate it. Absent: docs, llms_txt, agent_tooling_artifacts, quickstart, authentication, request_examples.

就绪度维度

评估维度得分
文档质量10
执行结果可验证性0
机器接口13
项目定位清晰度75
资源可发现性30
工作流完整度8

对 Agent 有帮助的部分

  • sitemap: verified during this run

Agent 受阻的部分

  • No documentation or developer pages discovered from the entry page or well-known paths.
  • llms.txt is absent (HTTP probe during this run).
  • No agent instruction files, code-distribution commands, or named slash-command skills found across fetched pages.
  • No quickstart signal matched across 1 fetched pages.
  • No authentication signal matched across 1 fetched pages.
  • No request examples signal matched across 1 fetched pages.

证据核查

关于该工具的公开声明，每条均标注核验状态与引用来源。

blog-agent-security-classification-in-agent-era.html4
dknownai.com已验证核验于 2026年7月15日

The four risk classifications are AGENT_HACK, SYS_FLAG, CONTENT_FLAG, and SAFE.

Many traditional AI safety rules do not translate cleanly into Agent workflows because agents interact with files, tools, APIs, and system workflows.

Developers can integrate DKnownAI Guard into agent tool chains using Guardrail-Bridge.

The system distinguishes between attacks and legitimate operations — for example, reviewing an API key setup flow to prevent hardcoding secrets is not treated the same as stealing credentials.

https://dknownai.com/blog-agent-security-classification-in-agent-era.html
docs.html3
dknownai.com已验证核验于 2026年7月15日

The API returns one of four risk classifications for each input, distinguishing deceptive tactics from direct high-risk requests.

DKnownAI Guard exposes a single REST endpoint requiring one required field and returning clear results.

The API is accessed via POST to https://open.dknownai.com/v1/guard with an API key in the Authorization header.

https://dknownai.com/docs.html
AI Agent Security API | Prompt Injection Detection | DKnownAI2
dknownai.com已验证核验于 2026年8月30日

The entry page was fetched and analyzed for machine-interface signals (title, headings, developer links, keyword probes).

Agent-native positioning as a marketing claim without a documented path: "The page mentions being built for agents like Claude Code but does not provide a concrete operational path for agents (e.g., AGENTS.md, slash commands, or agent-specific quick start).".

https://dknownai.com/
index.html2
dknownai.com已验证核验于 2026年7月15日

DKnownAI Guard is a security API built for agentic AI that detects manipulation and prompt injection.

DKnownAI Guard focuses on detecting whether someone is trying to manipulate the agent, separate from whether text content itself is harmful.

https://dknownai.com/index.html
blog-evaluating-ai-agent-security-guardrails.html2
dknownai.com已验证核验于 2026年7月15日

DKnownAI Guard achieved a 90.4% true negative rate in the vendor's evaluation, suggesting strong false-positive handling.

The system was evaluated against agent threats including instruction override, indirect injection, role hijacking, chain-of-thought poisoning, privacy leakage attempts, and tool abuse.

https://dknownai.com/blog-evaluating-ai-agent-security-guardrails.html
https://dknownai.com/sitemap.xml1
dknownai.com已验证核验于 2026年8月30日

sitemap.xml is reachable and lists site pages.

https://dknownai.com/sitemap.xml
blog-grok-agent-hack-dknownai-guard.html1
dknownai.com已验证核验于 2026年7月15日

A published case study showed Grok 4.5 being induced to organize high-risk chemical and biological content into structured, professional-looking documents.

https://dknownai.com/blog-grok-agent-hack-dknownai-guard.html

决策核对台

在依赖该产品或访问官网前，最值得先确认的问题。

DKnownAI Guard is a security API built for agentic AI that detects prompt injection and manipulation attempts, classifying input into four risk categories to help developers protect agent workflows from hijacking.

Traditional content moderation checks whether text itself is harmful. DKnownAI Guard checks whether someone is trying to manipulate the agent, separating deceptive tactics from content-level risk so legitimate requests are not overblocked.

It returns one of four classifications: AGENT_HACK (detected manipulation or jailbreak), SYS_FLAG (operational concern requiring scope review), CONTENT_FLAG (harmful content detected), or SAFE (no risk).

Make a POST request to open.dknownai.com/v1/guard with your API key in the Authorization header, or use Guardrail-Bridge for deeper integration into agent tool chains.

The vendor published a case study demonstrating detection of Grok 4.5 being induced to generate dangerous chemical and biological content, and reported a 90.4% true negative rate in their evaluation against agent-specific threats. Independent third-party verification is not yet available.

请在官网核验

继续探索

相近任务的不同路径

这些工具以带有明确编辑理由的替代关系关联到当前产品。

01ExtWise

ExtWise

AI developer tool in the same ecosystem addressing adjacent agent workflow challenges.

查看档案
02Claude Buddy

Claude Buddy

AI developer tool focused on agent assistance, complementary to security-focused tooling in the agent stack.

查看档案
03CodingPlan

CodingPlan

AI developer tool for coding workflows where prompt injection security considerations are directly relevant.

查看档案
查看 DKnownAI Guard 的全部替代工具