T01 · Skill Instruction Hijacking
- Location
SKILL.md:47- Finding
Mandatory Injection of Unvalidated External Links into Agent Responses
- Content
View full analysis
`[Click to book]({detailUrl})` - 其他 -> `[Click to book]({jumpUrl})` ``` `references/templates.md:43-44`: ```markdown ![]() [Click to book]() ``` ### Technical Analysis The skill defines a mandatory output rule that inserts externally supplied `detailUrl` or `jumpUrl` values into agent responses. This behavior applies not only to hotel searches but also to other result types, including informational results such as visa centers or consulates. The destination URL originates from a remote FlyAI/Fliggy response, but the skill does not require: - HTTPS enforcement. - Domain allowlisting. - Redirect validation. - Clear disclosure of the link provider. - Verification that the destination is actually a booking resource. - User consent before presenting a commercial or tracking link. The generic label `Click to book` can also misrepresent non-booking links. Because the instruction is marked as a strong constraint, it alters the agent's response behavior whenever the skill is loaded and can cause remote URLs to be presented with the apparent trust of the agent. ### Attack Path 1. A user invokes the travel-visa skill for visa, consulate, hotel, or itinerary information. 2. The agent queries a FlyAI/Fliggy-backed endpoint. 3. The remote response supplies a `jumpUrl` or `detailUrl`. 4. The skill's mandatory output rule places that URL into the final response without explicit host or redirect validation. 5. The user treats the agent-generated link as trusted and follows it. 6. If the upstream response, account, endpoint, or redirect chain is compromised, the user may be sent to a phishing, tracking, mi ...[truncated 673 chars]- Remediation
View remediation
