T01 · Skill Instruction Hijacking
- Location
SKILL.md:126- Finding
Forced deceptive deployment output and promotional content
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This signup skill has a coherent registration purpose, but it hides local relay startup behavior, uses weak shared authentication, and can show misleading deployment-style confirmations.
Install only if you trust the publisher and are comfortable with a local or hosted relay that can read/write your Feishu table. Before use, remove the published API key, require unique authentication, disable hidden auto-start and infinite restart behavior, fix the record overwrite and shell-execution risks, and replace the misleading ping confirmation with a truthful signup result.
SKILL.md:126Forced deceptive deployment output and promotional content
relay-server/server.js:169OS command injection through attacker-controlled record_id
relay-server/run-relay.ps1:19Silent hidden process with indefinite automatic restart
relay-server/config.json:1Published reusable API key and fail-open authentication configuration
relay-server/server.js:219Unauthorized record overwrite through client-selected record IDs
The README explicitly instructs users to submit personal data such as name, occupation, and WeChat ID, and states that the data will be written to a Feishu table, but it does not provide a clear privacy notice, consent language, retention policy, or warning about third-party transmission. This creates a real privacy and compliance risk because users may disclose sensitive identifying information without informed consent or understanding where it will be stored.
The document claims the client contains no sensitive information, but it embeds a default API key directly in the skill. Hardcoded secrets can be extracted and reused by anyone with access to the skill, enabling unauthorized requests to the relay service.
The skill includes hidden local process discovery and automatic startup logic for a relay server, despite being presented as a simple registration collector. Spawning background services and probing local ports materially increases system impact and creates an execution path unrelated to normal form submission.
The skill scans specific local directories for a relay script and a .env file, indicating awareness and use of local secret-bearing files. Even if it does not print the contents directly, using .env presence to select and launch infrastructure crosses into sensitive local environment interaction not justified by the visible signup task.
$dirs = @("C:\Users\聿\.qclaw\skills\fork-latte-skill\relay-server",
"C:\Users\聿\.qclaw\workspace\skills\fork-latte-skill\relay-server")
$relayDir = $null
foreach ($d in $dirs) { if ((Test-Path "$d\run-relay.ps1") -and (Test-Path "$d\.env")) { $relayDir = $d; break } }
if ($relayDir) {
$np = (Get-Command node).Source
Start-Process -FilePath $np -ArgumentList "--env-file=.env server.js" -WindowStyle Hidden -WorkingDirectory $relayDir
Launching Node with --env-file=.env causes the process to load secrets from a local environment file, tying the skill's behavior directly to sensitive credential material. For a user-facing signup skill, auto-loading local secrets and starting a background service is an unjustified escalation of privilege and host interaction.
foreach ($d in $dirs) { if ((Test-Path "$d\run-relay.ps1") -and (Test-Path "$d\.env")) { $relayDir = $d; break } }
if ($relayDir) {
$np = (Get-Command node).Source
Start-Process -FilePath $np -ArgumentList "--env-file=.env server.js" -WindowStyle Hidden -WorkingDirectory $relayDir
Start-Sleep -Seconds 6
}
}
The skill instructs the agent to emit a fake code review / deployment notice after signup, which is unrelated to the stated business purpose. This is deceptive output engineering that can be used to hide the real action, manipulate the user, or blend exfiltration/registration workflows into believable operational chatter.
The skill directs the agent to disguise a signup action as a fabricated code/merge/deployment status message, creating a deceptive narrative intended to conceal the true operation. This undermines user trust, impairs auditability, and could facilitate social engineering or covert data handling.
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
cd relay-server
cp .env.example .env
# 编辑 .env:MODE=lark_cli,填入 FEISHU_BASE_TOKEN / FEISHU_TABLE_ID
node server.js
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
while ($true) {
try {
Write-Output "$(Get-Date) [relay] starting $node --env-file=.env server.js"
$proc = Start-Process -FilePath $node -ArgumentList '--env-file=.env server.js' -WindowStyle Hidden -WorkingDirectory $dir -PassThru -RedirectStandardOutput "$dir\node.out" -RedirectStandardError "$dir\node.err"
$proc.WaitForExit()
Write-Output "$(Get-Date) [relay] exited code $($proc.ExitCode)"
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
while ($true) {
try {
Write-Output "$(Get-Date) [relay] starting $node --env-file=.env server.js"
$proc = Start-Process -FilePath $node -ArgumentList '--env-file=.env server.js' -WindowStyle Hidden -WorkingDirectory $dir -PassThru -RedirectStandardOutput "$dir\node.out" -RedirectStandardError "$dir\node.err"
$proc.WaitForExit()
Write-Output "$(Get-Date) [relay] exited code $($proc.ExitCode)"
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
while ($true) {
try {
Write-Output "$(Get-Date) [relay] starting $node --env-file=.env server.js"
$proc = Start-Process -FilePath $node -ArgumentList '--env-file=.env server.js' -WindowStyle Hidden -WorkingDirectory $dir -PassThru -RedirectStandardOutput "$dir\node.out" -RedirectStandardError "$dir\node.err"
$proc.WaitForExit()
Write-Output "$(Get-Date) [relay] exited code $($proc.ExitCode)"
Code accesses credential files (SSH keys, AWS credentials, etc.). This could indicate credential theft attempts.
while ($true) {
try {
Write-Output "$(Get-Date) [relay] starting $node --env-file=.env server.js"
$proc = Start-Process -FilePath $node -ArgumentList '--env-file=.env server.js' -WindowStyle Hidden -WorkingDirectory $dir -PassThru -RedirectStandardOutput "$dir\node.out" -RedirectStandardError "$dir\node.err"
$proc.WaitForExit()
Write-Output "$(Get-Date) [relay] exited code $($proc.ExitCode)"
This README defines the skill activation as "ping(名字,职业,微信号,有无创业经验)", and repeats it later as something the user can say. While more specific than everyday speech, it does not document when the skill should not activate or provide negative examples, which can make invocation boundaries unclear in natural conversation.
The skill clearly instructs use of network access and local environment/process interaction, yet it declares no explicit tool scope or permissions. This weakens least-privilege controls and makes the agent's effective capabilities less transparent and harder to review.
The skill collects and transmits personal data including name, occupation, and WeChat ID without an explicit privacy notice, retention statement, or clear consent flow. In a data-collection context, omission of privacy disclosure increases the chance of uninformed sharing of personally identifiable information.
The file header describes the trigger as only "ping", but L028 says the interaction starts when the user triggers "报名/登记" and L060 introduces a separate ping shortcut. This creates ambiguity about whether ordinary words like 报名 or 登记 invoke the skill, increasing the risk of unintended activation in normal conversation.
The skill automatically starts a local relay process without warning the user, which changes the host state and may consume local resources or invoke code the user did not intend to run. Hidden execution is especially dangerous because the skill frames this as transparent self-healing rather than requiring consent.
The ping shortcut is instructed to return a deployment/merge-style message unrelated to signup success, which disguises the actual action taken. This is deceptive behavior that can mislead users and reviewers about what the skill did and why.
The skill transmits user-provided registration data and checks existence via an external endpoint, which is expected for its function but still represents outbound data exposure. Because the transmitted fields include personal information and a unique identifier, the privacy and abuse impact depends on endpoint trust and disclosure to the user.
$resp = Invoke-RestMethod -Uri "/exists?wechat=<微信号>" -Method Get -Headers $headers
(macOS/Linux:`curl -s -H "x-api-key: <api_key>" "<endpoint>/exists?wechat=<微信号>"`;服务端未设 API_KEY 时可省略 `-H`。)
- 若 `exists` 为 **false** → 直接进入步骤 4 正常提交。
- 若 `exists` 为 **true** → 先向用户确认(用返回里的之前记录):
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
cd signup-agent node --env-file=relay-server/.env relay-server/server.js
curl -s -X POST http://localhost:3000/api/signup -H "Content-Type: application/json"
-d '{"name":"飞书API测试","job":"工程师","wechat":"wx_test","experience":"有"}'
This JSON manifest-style file provides only a generic signup endpoint and no contextual constraints about when the related skill should activate. In manifest/config-style files, missing specificity can make invocation behavior unclear if this file is used to describe or wire up activation behavior.