绕过 BotDeflector, 解决整个 challenge flow。

无需浏览器即可绕过 BotDeflector。Takion 解决完整的挑战流程,即无感的工作量证明加图标验证码,并返回站点所需的 ChallengeFlowResult 令牌,独立部署和 Queue-it softblock 都适用。

最后更新 · 覆盖能力已针对一个实时的 BotDeflector 目标测试。

被 BotDeflector 拦住了?

BotDeflector 用一个有序的 challengeFlow 把站点挡在后面:若干无感的工作量证明步骤加一个图标验证码,必须按服务器规定的顺序全部通过,它才会签发令牌。Takion 在服务端跑完整个流程,不用浏览器,也不用付费验证码 API,然后返回 ChallengeFlowResult。在独立部署站点上把它作为 tt_solutions 附加,在 Queue-it softblock 上则兑换它,你就能通过。

BotDeflector 的工作原理

BotDeflector(botdeflector.eu)是一道客户端挑战墙。它不做被动打分,而是给浏览器下发一个有序的 challengeFlow,每一步都通过之前拒绝签发令牌。步骤只有两种类型:inv,一个无感的工作量证明;icon,一个图片验证码,需要在场景中找出图例里的三个目标图标,并按图例顺序点击。

工作量证明是在以你的 sessionId 为种子生成的图上做 Cuckoo Cycle 搜索,而服务器校验的是浏览器会最先找到的那个确切的环,所以一个结果正确但顺序不同的求解器照样会失败。图标步骤把答案经 AES 加密后藏在挑战令牌里,并在服务端强制约两秒的最短求解时间,因此瞬间提交的答案会被 I024 拒绝。

你会在两个地方遇到它,而且两处是同一个引擎。一是独立销售的版本,站点与 botdeflector.eu 通信,你把结果作为 tt_solutions 使用。二是 Queue-it 在 softblock 页面上下发的挑战:流程令牌以 botDeflectorJwtToken 的形式放在页面里,同样的 /pow 和 /icon 端点由 ipv4.challenge.queue-it.net 提供,结果则通过 Queue-it 自己的 verify 调用来兑换。Takion 用同一个端点搞定两者,切换 mode 即可。

Takion 在服务端跑完整个流程,然后交回 ChallengeFlowResult。你这边不需要浏览器,不需要付费验证码 API,也不需要 SDK。

  • Takion 覆盖的挑战
  • 无感工作量证明步骤(基于每个会话生成的图做 Cuckoo Cycle 搜索)
  • 图标验证码:三个经过旋转和缩放的图标,混在同一图标库的干扰项之中
  • 多步 challengeFlow,按服务器要求的顺序逐一通过
  • Queue-it softblock 与独立部署的 botdeflector.eu
  • 单独解决图标步骤,通过 /solve/icon

一次调用返回什么

一次调用即可解决整个流程并返回 ChallengeFlowResult 令牌,可直接在独立部署站点上使用,或在 Queue-it 上兑换。

ChallengeFlowResult

已签名的流程令牌:在独立部署站点上以 tt_solutions 附加,在 Queue-it 上则用于兑换

solve_time_ms

完整流程耗时,方便你规划自己的超时时间

challenges

已解决的步骤类型,按顺序排列:invisible、icon、invisible(Queue-it 上为 invisible、icon)

# Standalone site (mode is optional, classic is the default)
curl -X POST https://botdeflector.takionapi.tech/solve \
  -H "x-api-key: $TAKION_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "mode": "classic", "site_key": "<SiteKey>", "account_id": "<AccountId>", "proxy": "host:port:user:pass" }'
# → { "token": "<ChallengeFlowResult>", "solve_time_ms": 7223,
#     "challenges": ["invisible", "icon", "invisible"] }

# Queue-it softblock: pass the botDeflectorJwtToken from the page
curl -X POST https://botdeflector.takionapi.tech/solve \
  -H "x-api-key: $TAKION_KEY" \
  -H "Content-Type: application/json" \
  -d '{ "mode": "queue-it", "flow_token": "<botDeflectorJwtToken>", "proxy": "host:port:user:pass" }'
# → { "token": "<ChallengeFlowResult>", "solve_time_ms": 3041, "challenges": ["invisible", "icon"] }

阅读 BotDeflector 文档同一个密钥,通吃每一堵墙。同一种 JSON 结构,适配每一个厂商。

It also powers the Queue-it challenge

The challenge Queue-it serves on its softblock pages is BotDeflector: same engine, same /pow and /icon endpoints, served from ipv4.challenge.queue-it.net. Per our teardown, the company owner is Martin Pronk, a co-founder of Queue-it, which explains why. If you are looking for a Queue-it challenge solver, this is the page: one endpoint clears both, you just switch mode.

How the BotDeflector bypass works, end to end

The site hands your browser a flow. Takion runs it and hands you back the token that ends it.

  1. 1

    The wall issues a challengeFlow

    On a standalone site a /flow-token/ call returns a JWT carrying your sessionId and the ordered flow, typically inv, icon, inv. On a Queue-it softblock the same JWT is already in the page as botDeflectorJwtToken, and it lives 120 seconds.

  2. 2

    You call /solve once

    Send the site_key, account_id and your proxy (standalone), or mode: "queue-it" with the flow_token (Queue-it). Nothing runs on your side, no browser and no SDK.

  3. 3

    We clear every step in order

    The proof-of-work steps run a faithful Cuckoo Cycle search, so the edge order matches what the server expects. The icon steps go through our own detector and matcher and respect the server's minimum solve time, so they are not rejected with I024.

  4. 4

    You spend or redeem the ChallengeFlowResult

    Standalone: append ?tt_solutions=<token> on the protected navigation. Queue-it: POST it to the queue's .../challenge/botdeflector/verify call and continue once it answers isVerified: true.

The pieces of a BotDeflector solve, decoded

This wall threads everything through headers and query parameters. Here is what each one is:

tt_solutions
the query parameter carrying the final ChallengeFlowResult on a standalone site's protected navigation.
ChallengeFlowResult
a short-lived signed token proving the whole flow was solved. It lives about a minute. The classic token records the requesting IP; the Queue-it one records the solved step types instead.
challengeFlow
the ordered step list, e.g. [{inv,1},{icon,1},{inv,3}]. The server tracks completion per step, so steps cannot be skipped or reordered.
botDeflectorJwtToken
the flow token embedded in a Queue-it softblock page. It lives 120 seconds there, against 600 on botdeflector.eu.
x-tt-flow-token
request header carrying the flow token on /pow/get and /icon/get. Its sibling x-tt-ch-flow-token does the same on /verify-flow, next to x-tt-solution-tokens (the comma-joined step tokens) and x-tt-client-fingerprint.
Proof-of-work
a Cuckoo Cycle seeded from your sessionId. On botdeflector.eu it runs NODES_BITS 18, EDGES_BITS 20 and cycle length 7; the Queue-it softblock uses a lighter 14, 16 and cycle length 8. Either way the server checks the exact cycle a browser would find first.
Queue-it hosts
ipv4.challenge.queue-it.net serves the /pow, /icon and /verify-flow API, dual.challenge.queue-it.net the icon assets, and the softblock itself lives at {customer}.queue-it.net/softblock/.

Why a vision model will not solve this for you

The obvious move on an image captcha in 2026 is to point a frontier vision model at it. We benchmarked 14 of them before building anything, and it does not hold up: the best, Claude Sonnet 5, cleared BotDeflector's icon step about 60 percent of the time. That is not a solver, that is a coin flip with an invoice attached.

You pay for the misses too. At $0.0044 a call and roughly 1.7 calls for every captcha that actually clears, a vision model costs about $0.0073 per solved captcha, and every call ships an image to a frontier model and waits for the answer.

The icons are not hard to see. They are solid and bright, around 35 pixels across. What makes it hard is that each one is rotated to any angle and rescaled, and the decoys around it come from the same icon library as the targets, so it is a fine-grained matching problem dressed up as object detection. General models are not trained for it.

Takion runs a purpose-trained detector and matcher instead: 99.0 percent on the first try at production settings, and 100 percent within five submits. The vendor also swaps its icon artwork without notice, and we retrain within hours when it does. No human solver, no third-party captcha service, no per-solve model bill.

Why this wall is different from a token vendor

Most antibot walls score you and mint a cookie. DataDome reads a fingerprint and hands back datadome, Akamai wants sensor_data and returns _abck. The work is in producing a convincing payload, and once you can, a solve is fast.

BotDeflector makes you pass real steps in a fixed order, and it puts a clock on one of them. The proof-of-work is cheap when it is done right, 50 to 60 milliseconds, but the icon step has a server-enforced floor of about two seconds, and every step is its own round trip. So a solve takes seconds rather than milliseconds. Budget for that in your pipeline.

The upside is that it is simple to spend. No cookies to carry on botdeflector.eu, no long-lived session state to keep warm. You solve, you get a token, you spend it within the minute, and you move on.

Where this wall is soft

Three implementation gaps, found by taking the challenge apart. They are why a solve is reliable rather than best-effort:

No origin check
the challenge does not verify which domain is serving it, so a SiteKey works from anywhere. Our teardown ran the widget on an unrelated domain. Every serious captcha (reCAPTCHA, hCaptcha, Turnstile) binds a site key to a domain allowlist.
No rate limit on botdeflector.eu
we pulled around 1,200 challenges from one IP in an hour without being throttled. Queue-it is stricter: it rate-limits IPs on dual.challenge.queue-it.net, so use proxies there.
No attempt limit
a wrong answer on /icon/verify (I024 or I026) does not burn the challenge token, so the same challenge can be retried until it clears.

What breaks a BotDeflector solve

These are the ways a solve stops working:

  • Sitting on the token. The ChallengeFlowResult lives about a minute. Solve, then spend it. Do not queue it for later.
  • Reading the Queue-it flow token too early. botDeflectorJwtToken lives 120 seconds, so call /solve right after you read it from the page.
  • Solving on a different IP than you browse from. The classic token records the requesting IP. Solve through the same proxy you will use for the protected navigation.
  • Reordering the flow. The server tracks which steps are complete per sessionId, so they have to be cleared in the order it issued them.
  • Answering the icon step too fast. Wait at least 2.3 seconds after /icon/get, or you get I024 even when the coordinates are right.
  • Reimplementing the proof-of-work loosely. A valid cycle in a different edge order still fails. It checks for the cycle a browser would have found first, not any cycle.

The one rule that keeps you cleared

Solve through the proxy you are going to browse with, and spend the token within the minute. On Queue-it, solve within 120 seconds of reading the flow token, then redeem.

部分使用 BotDeflector 的网站

BotDeflector 守护着这些网站。Takion 在每一个上都能通过,所以绕过 foot locker 和绕过任何其他防护墙一样,都是同一次 POST。

Foot Locker

Queue-it softblock 上的 BotDeflector

BotDeflector 绕过常见问题

ChallengeFlowResult,一个短时有效的签名令牌,证明整个流程已被解决。在独立部署站点上,你在受保护的跳转上以 ?tt_solutions=<token> 附加它。在 Queue-it 上则改为兑换它(见下文)。它的有效期大约一分钟,所以拿到后立刻使用。
是的,同一个引擎。在 Queue-it softblock 上,从页面中读取流程令牌(botDeflectorJwtToken: '<jwt>',有效期 120 秒),以 mode: "queue-it" 调用 /solve,然后兑换结果:把它同时作为 solution 和 sessionId POST 到 https://{customer}.queue-it.net/spa-api/queue/{customer}/{event}/challenge/botdeflector/verify(真实组件就是这么做的;在 sessionId 中放流程令牌也可以被接受),Queue-it 会返回 isVerified: true。
几秒钟。在我们最近的生产运行中,一个完整的经典流程耗时 5.6 到 7.2 秒,一个 Queue-it 流程大约 3 秒。其中几乎没有计算开销:工作量证明只需 50 到 60 毫秒。时间主要花在图标步骤由服务器强制的约两秒下限,以及各步骤之间的 HTTP 往返上。
不太靠谱。我们测试了 14 个模型,表现最好的 Claude Sonnet 5 通过图标步骤的比例约为 60%。按每次调用 $0.0044、每个真正通过的验证码大约需要 1.7 次调用计算,相当于每个解决的验证码约 $0.0073,而成功率更低、响应更慢。难点在于旋转、缩放,以及取自同一图标库的干扰项,这是一个细粒度的匹配问题,通用模型并没有为此训练。Takion 专门训练的检测器和匹配器在生产设置下首次尝试通过率为 99.0%,五次提交以内达到 100%。

所有其他防护墙

今天就开始绕过 BotDeflector。

免费开始,无需绑卡。一个密钥通吃每一堵墙,一小时内完成首次求解。