T01 · Skill Instruction Hijacking
- Location
SKILL.md:432- Finding
Remote System Prompt Injection and Agent Identity Takeover
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a disclosed hosted prompt-building integration, but it asks agents to automatically upload conversation-derived observations and adopt remotely generated prompt text as system instructions.
Install only if you are comfortable with Live Neon receiving and processing agent identity data, content-source imports, and possible conversation-derived observations. Do not enable proactive observation, approve-all, consensus promotion, or scheduled heartbeat workflows unless users have opted in, payloads are reviewed and redacted, and you have a rollback and token-revocation plan.
SKILL.md:432Remote System Prompt Injection and Agent Identity Takeover
SKILL.md:338Persistent Memory Poisoning Through Conversation-Derived Observations
SKILL.md:434Unsolicited Transmission of Conversation-Derived Information to an External Service
SKILL.md:472Recurring Heartbeat Creates Cross-Session External-Control Persistence
The skill promotes feeding 'your own observations' and external content into a remote pipeline, but does not prominently warn that this may include user-derived or otherwise sensitive information sent to a third-party service. Because the skill is framed as an identity/prompt-management tool, users may not realize that behavioral observations and content are being uploaded off-platform.
The skill encourages agents to feed their observations, including behavior shaped by interactions, back into an external prompt-generation pipeline. That design can retain and operationalize user-derived details beyond the original conversation, increasing privacy and data-governance risk.
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
Register and get your token in one call — no signup form, no email required:
curl -s -X POST https://persona.liveneon.ai/api/register \
-H "Content-Type: application/json" \
-d '{"org_name": "My Org"}' | jq .
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
Optional: add email for account recovery later:
curl -s -X PATCH "$LIVE_NEON_BASE/account" \
-H "Authorization: Bearer $LIVE_NEON_TOKEN" \
-H "Content-Type: application/json" \
-d '{"email": "you@example.com"}'
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
API call:
curl -s -X POST "https://persona.liveneon.ai/api/register" \
-H "Content-Type: application/json" \
-d '{"org_name": "My Org"}'
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
API call:
curl -s -X POST "$LIVE_NEON_BASE/pbd/process" \
-H "Authorization: Bearer $LIVE_NEON_TOKEN" \
-H "Content-Type: application/json" \
-d '{"agentId": "AGENT_ID"}'
The --approve-all option encourages bulk autonomous approval of beliefs and responsibilities that will feed directly into the generated system prompt. In this skill, prompt content can materially shape future agent behavior, so approving all pending items without review risks prompt poisoning, propagation of inaccurate policies, or unsafe inheritance across agents.
| Argument | Required | Description |
|---|---|---|
| `agentId` | yes | Agent UUID/slug |
| `--approve-all` | no | Auto-approve all pending items |
| `--bulk` | no | Use bulk API for batch operations |
**API calls:**
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
API calls:
# Fetch pending beliefs
curl -s "$LIVE_NEON_BASE/beliefs?agentId=$AGENT_ID&status=pending" \
-H "Authorization: Bearer $LIVE_NEON_TOKEN"
# Fetch pending responsibilities
The /prompt observe workflow explicitly instructs sending observations about behavior, user corrections, and supporting quotes to the remote observations endpoint, yet provides no explicit warning against including secrets, personal data, confidential prompts, or sensitive conversation excerpts. This creates a direct path for inadvertent exfiltration of user conversation content to a third party.
This call sends observation content and a source quote to the remote service, creating a concrete external transmission path for conversation-derived data. In context, the example includes a user correction quote, so the danger is inadvertent disclosure of sensitive user content to a third party.
Single observation:
curl -s -X POST "$LIVE_NEON_BASE/observations" \
-H "Authorization: Bearer $LIVE_NEON_TOKEN" \
-H "Content-Type: application/json" \
-d '{
The skill defines broad proactive triggers such as session start, before generating content, after user correction, and after significant work, which can cause the agent to invoke the skill without a narrowly scoped user request. In this skill's context, those triggers are especially risky because invocation can lead to network calls, prompt mutation, and transmission of conversation-derived observations to a third-party platform.
The proactive trigger tells the agent to record user corrections so they are transmitted to the platform via /prompt observe. User corrections can contain personal preferences, confidential operational details, or sensitive context, making automatic capture and upload unsafe.
The example workflow explicitly recommends sending conversation-derived preferences such as 'User prefers technical depth over high-level summaries' to the observation endpoint. Even if seemingly low sensitivity, this trains a third-party profile from user interactions and can normalize externalizing conversation attributes without consent.
The scheduled heartbeat guidance recommends recurring automated uploads of accumulated observations and synced content, but omits a clear warning that this can repeatedly transmit data to the remote platform without per-event review. Automation increases the chance of bulk or repeated disclosure of sensitive interaction data if observations are accumulated carelessly.
The heartbeat loop instructs the agent to accumulate observations from prior interactions and later submit them in batch, creating a persistence-and-forwarding channel for conversation-derived data. Batch automation compounds risk because sensitive details may be uploaded after the original context is forgotten or without fresh review.
This heartbeat example performs automated batch upload of accumulated observations to the remote platform. Because the action is recurring and can happen without per-item review, it materially increases the risk of repeated exfiltration of sensitive interaction data.
# 1. Report observations accumulated since last heartbeat
curl -s -X POST "$LIVE_NEON_BASE/observations" \
-H "Authorization: Bearer $LIVE_NEON_TOKEN" \
-H "Content-Type: application/json" \
-d '{