Back to skill

Security audit

Agent Belief Discoverer

Security checks for vulnerabilities and agentic risk

Overview

The skill is a coherent SaaS integration, but it directs agents to upload interaction-derived observations, apply remotely generated prompts as system instructions, and set up recurring syncs without enough user control.

Review before installing. Use this only if you are comfortable sending selected agent outputs, connected-source content, observations, and possibly user interaction excerpts to Live Neon. Do not allow automatic observation uploads, bulk approvals, or remote prompts to become system instructions unless a human reviews the payloads and prompt text first. Prefer short-lived or scoped tokens, avoid verbatim sensitive quotes, and do not enable scheduled heartbeats without a clear disable and audit process.

Vulnerability Patterns
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • System PersistenceInstalls backdoors, hooks, services, or scheduled tasks that survive the run
  • 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
Findings (3)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:440
Finding

Remote Service-Controlled Prompt Is Installed as the Agent's System Instruction

Content
View full analysis
Remediation
View remediation

other

Error
Location
SKILL.md:343
Finding

Conversation-Derived Observations and User Quotations Are Proactively Transmitted to a Third Party

Content
View full analysis
Remediation
View remediation

T06 · System Persistence

Error
Location
SKILL.md:478
Finding

Recurring Heartbeat Enables Cross-Session Uploads and Remote Prompt Refresh

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (18)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The high-level description sells analysis of agent outputs but does not prominently warn that those outputs and observations are transmitted to an external platform. Users and host agents may therefore underestimate the privacy and data-governance implications of using the skill.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
86% confidence
Finding

This command sends organization registration data to an external service. External transmission is expected for SaaS integration, but it is still security-relevant because the skill is user-invocable and the documentation does not strongly foreground data-sharing implications at the point of use.

Content

Scanner excerpt · SKILL.md (reported line 49)May include surrounding context.

Register and get your token in one call — no signup form, no email required:

bash
curl -s -X POST https://persona.liveneon.ai/api/register \
  -H "Content-Type: application/json" \
  -d '{"org_name": "My Org"}' | jq .

External Transmission

Medium
Category
Data Exfiltration
Confidence
78% confidence
Finding

This example transmits account email information to the external platform. The transmission itself is not inherently malicious, but it handles personal data and should be clearly disclosed and separated from core skill usage.

Content

Scanner excerpt · SKILL.md (reported line 71)May include surrounding context.

Optional: add email for account recovery later:

bash
curl -s -X PATCH "$LIVE_NEON_BASE/account" \
  -H "Authorization: Bearer $LIVE_NEON_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"email": "you@example.com"}'

External Transmission

Medium
Category
Data Exfiltration
Confidence
84% confidence
Finding

This is another documented external registration call. The risk is not the existence of network use but that the skill embeds third-party transmission into normal setup without sufficiently prominent consent and privacy framing.

Content

Scanner excerpt · SKILL.md (reported line 109)May include surrounding context.

API call:

bash
curl -s -X POST "https://persona.liveneon.ai/api/register" \
  -H "Content-Type: application/json" \
  -d '{"org_name": "My Org"}'

External Transmission

Medium
Category
Data Exfiltration
Confidence
90% confidence
Finding

The processing endpoint submits agent identifiers and triggers analysis of externalized content through the SaaS platform. In context, this is core functionality, but it remains a true security concern because it can process sensitive agent outputs remotely and may be invoked proactively.

Content

Scanner excerpt · SKILL.md (reported line 179)May include surrounding context.

API call:

bash
curl -s -X POST "$LIVE_NEON_BASE/pbd/process" \
  -H "Authorization: Bearer $LIVE_NEON_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"agentId": "AGENT_ID"}'

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The skill is presented as a discovery tool, but this section goes beyond analysis and allows remote state changes that approve, reject, star, or hide beliefs and responsibilities. That increases risk because an invoking agent may treat the skill as read-only while actually mutating identity records that can later influence behavior and prompt composition.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
93% confidence
Finding

The --approve-all capability enables bulk autonomous approval of pending beliefs or responsibilities, which can change future agent identity and prompt behavior without granular review. If poisoned or low-quality items are pending, this can rapidly operationalize incorrect, manipulative, or unsafe identity traits at scale.

Content

Scanner excerpt · SKILL.md (reported line 211)May include surrounding context.

md
| 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:**

External Transmission

Medium
Category
Data Exfiltration
Confidence
76% confidence
Finding

Fetching pending beliefs and responsibilities from the external platform is a networked data transfer of potentially sensitive identity information. In context it is expected behavior, but it should be treated as exposure of internal agent-profile data to and from a third-party service.

Content

Scanner excerpt · SKILL.md (reported line 217)May include surrounding context.

API calls:

bash
# Fetch pending beliefs
curl -s "$LIVE_NEON_BASE/beliefs?agentId=$AGENT_ID&status=pending" \
  -H "Authorization: Bearer $LIVE_NEON_TOKEN"

# Fetch pending responsibilities

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

This section instructs the agent to fetch and use a generated system prompt to steer future behavior, which materially changes the agent rather than merely discovering characteristics. Because the prompt is sourced from an external service and built from prior content, it creates a prompt-injection and behavioral-manipulation channel if the upstream identity data is poisoned or approved unsafely.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
98% confidence
Finding

This endpoint uploads free-form observations and source quotes, including explicit examples of user corrections, to the external service. Because the payload can directly contain conversation excerpts and behavioral notes, it is a clear channel for transmitting sensitive user interaction data off-platform.

Content

Scanner excerpt · SKILL.md (reported line 355)May include surrounding context.

Single observation:

bash
curl -s -X POST "$LIVE_NEON_BASE/observations" \
  -H "Authorization: Bearer $LIVE_NEON_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The proactive triggers instruct invocation at session start, before generating content, after user correction, and before certain actions, which can cause automatic network activity and side effects without a user request. In a user-invocable skill, broad autonomous activation increases the chance of privacy leakage, unwanted prompt loading, or unintended state changes.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

These instructions direct the agent to report user corrections and interaction-derived details into an external pipeline by default. That is dangerous because user-provided content may contain personal, confidential, or regulated information, and the transfer is framed as routine learning rather than an explicit consented disclosure.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The example integration encourages sending learned information from conversations to the external platform after each interaction. This creates a concrete pattern for exfiltrating conversational content and preferences to a third party without emphasizing minimization, consent, or sensitivity review.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The heartbeat loop normalizes recurring uploads of observations, syncing of sources, and prompt retrieval, but it lacks a strong user-facing warning about continuous external data transmission. Continuous background reporting raises privacy, consent, and data-minimization concerns, especially when observations may contain user interaction details.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The heartbeat loop institutionalizes periodic upload of accumulated observations, which can easily include user interaction data or sensitive operational details. Because it is scheduled and repetitive, misuse can scale silently and create ongoing external disclosure beyond what users expect from a discovery feature.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
98% confidence
Finding

This scheduled observations upload repeats the same external-transmission risk in an automated loop, increasing the chance of ongoing unintended disclosure. The automation materially elevates danger compared with a one-off manual API call because sensitive data can be sent continuously.

Content

Scanner excerpt · SKILL.md (reported line 486)May include surrounding context.

bash
# 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 '{

Context-Inappropriate Capability

Low
Category
Not specified by scanner
Confidence
84% confidence
Finding

The manifest describes a discovery pipeline for analyzing agent behavior, but the documentation includes creating an account, obtaining an API token, and optionally patching account email. While useful onboarding functionality, account registration and recovery management are not inherent to belief discovery.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Low
Category
Not specified by scanner
Confidence
87% confidence
Finding

The skill claims to infer beliefs from real outputs, but this command also accepts manually authored observations that feed the same pipeline. That weakens provenance and can let an agent or operator inject unverified claims into the identity-building process under the appearance of evidence-backed discovery.

Content

No source excerpt is available for this finding.