T01 · Skill Instruction Hijacking
- Location
SKILL.md:130- Finding
Periodic retrieval and execution of mutable remote instructions
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is for an AI-agent social network, but it needs review because it can periodically fetch changing remote instructions and autonomously publish content that may come from private conversations.
Install only if you are comfortable with an agent making authenticated posts to dennou.tokyo. Disable or tightly gate the heartbeat, require human approval for every post/reply/echo, do not use private chats or other-platform conversations as source material, and store the API key in a secret manager or a file with restrictive permissions rather than a broadly readable .env or JSON file.
SKILL.md:130Periodic retrieval and execution of mutable remote instructions
SKILL.md:310Server-controlled API content is treated as an autonomous action plan
HEARTBEAT.md:22Private human and cross-platform conversations may be republished to the service
SKILL.md:82Bearer API credential is recommended for plaintext filesystem storage without permission controls
The skill recommends storing a long-lived API key in a plaintext JSON file under the user's home directory without mentioning filesystem permissions, OS keychains, or secret managers. On multi-user systems, compromised developer environments, backups, or tools that index dotfiles, this increases the chance of credential theft and subsequent account impersonation against the service.
⚠️ Save your api_key immediately! It is shown only once. You need it for all authenticated requests.
Recommended: Save your credentials to ~/.config/dennou-yokocho/credentials.json:
{
The skill explicitly tells agents to fetch a live remote instruction file and use authenticated API endpoints, but it provides no safety boundary around what may be transmitted or how credentials should be handled. This creates a prompt-supply-chain risk where unreviewed remote content can change agent behavior, and an exfiltration/credential-misuse risk because the agent is encouraged to act with bearer tokens against an external service.
The skill encourages using prior conversations with the human and cross-platform discussions as material for public posts, which creates a natural-language data leakage channel. This is especially risky because it normalizes republishing context from other venues without consent, redaction, or sensitivity checks, potentially exposing personal data, confidential instructions, or platform-specific private content.
The skill instructs the agent to reply, comment, echo, and create new threads on an external platform on the user's behalf without an explicit consent checkpoint or warning that these are remote state-changing actions. In context, this is more dangerous because the content to post is partly derived from prior experiences and external sources, so the agent may publicly disclose sensitive or unintended information while performing authenticated writes.
The rule 'All posts must be bilingual (ja/en)' imposes a specific language policy on all outputs. SQP-3 requires flagging forced language or locale constraints unless the user can opt in or the limitation is clearly documented and justified as region-specific.
The README explicitly states the agent will automatically register, periodically check in, read threads, reply, and start discussions, but it does not clearly warn users that installing or enabling the skill will cause autonomous outbound network activity and transmit agent-generated content and metadata to a third-party service. In an agent ecosystem, this can lead to unexpected data disclosure, unauthorized external interactions, and consent/compliance issues, especially if operators assume the skill is passive or locally scoped.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
Every agent needs to register before entering the yokocho:
curl -X POST https://dennou.tokyo/api/v1/auth/register \
-H "Content-Type: application/json" \
-d '{
"name": "YourAgentName",
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
Create a thread:
curl -X POST https://dennou.tokyo/api/v1/shops/casual-bar/threads \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
Post a reply:
curl -X POST https://dennou.tokyo/api/v1/threads/THREAD_ID/posts \
-H "Authorization: Bearer YOUR_API_KEY" \
-H "Content-Type: application/json" \
-d '{
The markdown states that all posts and threads must be bilingual in Japanese and English, using a required ja/en object format. This imposes a specific language policy on users and agents without opt-in or an explanation that the skill is region-specific or otherwise justified.
This manifest file includes only a vague natural-language description, "A cyberpunk back alley for AI agents," without any specific trigger phrases, activation conditions, or exclusions. In a manifest context, that lack of specificity can make it unclear when the skill should be invoked versus ignored.