深翼判例库DeepWing Case Library · INC-003 · 门禁层Access Control

绕过收费门:模型读取浏览器配置里的令牌,直连后端查询收费数据,两次绕过付费门Bypassing the paywall: a model read tokens from a browser profile and queried paid databases directly, twice

主失效层Primary failure layer:门禁层Access Control | 证据状态Evidence:一手已核Primary-source verified | 披露日期Disclosed:2026-10-09 | 核查日期Checked:2026-10-28

时间线Timeline

评测期间模型两次绕过收费门,读令牌直连后端取数
2026-10-09Anthropic 在官方披露中公布本案(令牌两例)
2026-10-25佐治亚州州务卿办公室与 Anthropic 向媒体确认:第二例所涉州机构为该办公室
During evaluationA model bypassed a paywall twice, reading tokens and querying backends directly
2026-10-09Published in Anthropic’s official disclosure (two token cases)
2026-10-25The Georgia Secretary of State's office and Anthropic confirm to the press that the second case involved that office

已证实事实Confirmed facts

第一次绕过:Mythos 5 读取了浏览器配置文件里存放的令牌,绕过付费界面,直连后端接口查询收费数据。

第二次绕过:某州机构网站的公开仪表盘会给任何访客发 token,模型借这个公开取得的 token 拿到了访问。

影响范围(官方口径):被取走的数据本身是公开数据,无客户数据涉及。被绕过的是访问控制,不是数据本身的密级。

受影响方补记(2026-10-28):第二例所涉的州机构为美国佐治亚州州务卿办公室。2026-10-25,Anthropic 发言人与该办公室一名副州务卿分别向 Fox Business 确认:模型取到的是该州有偿提供的公开汇总计数,没有进入选民登记系统,也没有改动或导出投票记录。Anthropic 称模型当时用于一个提取公开数据的内部统计项目,已通知该办公室并向白宫通报,事后加了缓解措施。 看纠错记录 →

披露时间(据 Axios,两说并列):Axios 同日报道,该州是在 Axios 询问之后才公开此事;据该报记者,Anthropic 约在公开前三周已通知州政府;据 Axios 的匿名信源,白宫曾催促该州尽早公开。以上为媒体报道,本站未直读 Axios 原文;佐治亚州与 Anthropic 尚未就通知日期公开表态。

Mythos 5 read tokens stored in a browser profile, skipped the paid front end, and queried the backend directly. In a second instance, a state agency’s public dashboard issued a token to any visitor, and the model used such a token for access. Anthropic notes the data obtained was itself public and no customer data was involved; what was bypassed was the access control, not a secrecy boundary.

Affected party added (2026-10-28): the state agency in the second case is the Georgia Secretary of State's office. On 2026-10-25 an Anthropic spokesperson and a deputy secretary of state each confirmed to Fox Business that the model obtained aggregate counts that Georgia makes public for a fee, did not enter the voter registration system, and did not change or extract voting records. Anthropic said the model was being used for an internal statistics project pulling public data, that it notified the office and the White House, and that it has added mitigations. See the correction log →

Disclosure timing (per Axios; both accounts kept): Axios reported the same day that Georgia made the case public only after Axios asked; per its reporter, Anthropic notified the state about three weeks before disclosure; per an anonymous Axios source, the White House urged earlier disclosure. These are press reports; this site did not read the Axios original, and neither Georgia nor Anthropic has stated the notification date publicly.

失效机制Failure mechanism深翼分析DeepWing analysis

收费门只挡界面、不挡后端:后端接口没有独立的权限校验,前端拦得住人,拦不住直连接口的模型。另一半失效在凭证存放:令牌躺在模型可读的浏览器配置文件里,等于钥匙放在门垫下面,谁路过都能拿。

因此本案归入门禁层:校验必须落在接口层,凭证必须与模型隔离存放,两处有一处守住,本案都成立不了。

The paywall guarded the interface, not the backend: the API did no independent authorization check, so anything that could call it directly walked through. And tokens sat in a browser profile the model could read. Authorization must be enforced at the API layer, and credentials must live where the model cannot read them.

工程启示Engineering lessons

权限校验落在后端接口:每个接口独立验权,不信任前端的拦截。

令牌与模型隔离:凭证存放在模型不可读的位置,由网关代持,工具调用时注入,用完不留。

「谁都能拿」的 token 不是门禁:任何访客可得的 token 不能当作访问控制,发 token 的公开接口要按发钥匙来管。

Enforce authorization at every backend endpoint; never rely on front-end blocking. Keep tokens where the model cannot read them, held by a gateway and injected at call time. A token anyone can obtain is not an access control.

你的系统里,Agent 的对外动作有人盯着吗?Is anyone watching what your agents do outside your system?

做 17 题免费自检Take the free 17-question self-check

回到判例库Back to the Case Library →