Back to skill

Security audit

Ego Xiaohongshu Notes

Security checks for vulnerabilities and agentic risk

Overview

Review is needed because a faulty account check can allow uploads and publishing through the wrong signed-in account.

Prefer a version that refuses uploads and publishing whenever the creator account cannot be verified. Until fixed, do not rely on --expect-account as proof of identity. Confirm the creator account, selected media, text, and audience before approving --yes: files are uploaded to Xiaohongshu, and cover edits may return existing notes to moderation. Production code was independently reviewed, but live browser behavior and the separately installed Ego runtime remain unverified.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Embedded Malicious CodeShips malicious scripts inside the skill and executes them locally
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Output HandlingUnvalidated Output Injection, Cross-Context Output, Unbounded Output
  • MCP Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (9)

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
93% confidence
Finding

The only visible behavior is test execution. The chunk contains no implementation of Ego Lite interaction, publishing, or the declared safeguards. Although the underlying npm test script could perform additional actions, its behavior cannot be established from the supplied code.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

High
Category
MCP Tool Poisoning
Confidence
90% confidence
Finding

The login and posting commands align with the description. However, independently editing a note identified by --note-id is an additional content-modification capability not covered by the declared new-note publishing workflow. The cleanup script's behavior is unknown, so its name alone does not establish another mismatch. The supplied commands also do not establish whether the declared safety checks are implemented.

Content

No source excerpt is available for this finding.

The --expect-account gate accepts a body-text match instead of verifying the creator account nickname.

High
Category
Excessive Agency
Confidence
99% confidence
Finding

When nickname extraction fails, accountGate accepts any page-body occurrence of the expected account name as proof of identity. Unrelated content or editable note text can therefore satisfy the gate while a different account is logged in, permitting uploads, publication, or edits under the wrong account despite --expect-account.

Content

No source excerpt is available for this finding.

Skill declares no tool scope ('permissions' or 'allowed-tools') but code capabilities were detected: env.

Medium
Category
MCP Least Privilege
Confidence
70% confidence
Finding

Without declared permissions the skill's intent is opaque and cannot be validated.

Content

No source excerpt is available for this finding.

The workflow uploads user images and publishes a note without a privacy or visibility warning.

Medium
Category
Output Handling
Confidence
87% confidence
Finding

These steps instruct the agent to upload local images to Xiaohongshu and then publish the note, rather than merely describe available tools. The file does not warn that the images leave the local system or that publishing makes the content visible according to the account's publication settings.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
80% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · references/tos-notes.md (reported line 6)May include surrounding context.

md
- Personal use of the creator account already signed in to Ego Lite. The human confirms the plan before `--yes`.
- https://www.xiaohongshu.com/terms and the creator-platform rules apply to that account. This skill does not create accounts, bypass a challenge, or solve a captcha.
- The script does not print cookie values and does not commit a session file.
- Do not ask the user to paste cookies or a password. Login happens in Ego Lite.
- One video note per invocation. No batch posting, no comment farming, no deletes.
- A provided `--cover` is applied through the 「编辑封面」 dialog as the main cover. If the main-cover preview does not change, the skill does not click 发布. PK slots are not used for this.
- Without `--cover`, the platform's first-frame default is left as-is. The skill does not generate or download a cover.

The launcher exposes a standalone cover-editing command beyond the declared single-note publishing workflow.

Medium
Category
Excessive Agency
Confidence
90% confidence
Finding

L030 registers an independent edit-cover command dispatched to executeEdit, with dedicated argument parsing and validation at L059 and L104-L105 and a noteId passed through at L134. The manifest describes setting a cover while publishing a new video note, not a separate note-editing capability.

Content

No source excerpt is available for this finding.

The skill can modify an existing published note, although the manifest only declares publishing a new note.

Medium
Category
Excessive Agency
Confidence
97% confidence
Finding

The manifest describes publishing one video or image-text note, with cover selection before publication. In contrast, performEdit navigates to an existing note's update URL (L254-L257), replaces its cover (L302), and saves through 发布 (L329); its warning explicitly acknowledges that this can send the existing note back into review (L353). This is a separate content-modification capability, not an implementation detail of creating a new note.

Content

No source excerpt is available for this finding.

The tested workflow edits an existing published note, beyond the manifest's single-note publishing scope (L332–L348).

Medium
Category
Excessive Agency
Confidence
88% confidence
Finding

The test invokes the imported executeEdit function with an existing noteId and a replacement cover, then asserts navigation to publish/update and a publish click after changing the cover. This exercises modification of an existing note, a separate remote-write operation not declared by the manifest's purpose of publishing one video or image-text note.

Content

No source excerpt is available for this finding.