Back to skill

Security audit

Agent Identity Evolution

Security checks for vulnerabilities and agentic risk

Overview

The skill is a coherent Live Neon integration, but it encourages ongoing third-party uploads and remotely generated identity prompts that can affect future agent behavior without strong consent or review gates.

Review this skill carefully before installing. Use it only for agents and content sources where you are comfortable sending behavioral observations, user corrections, quotes, and connected-source material to Live Neon. Do not let it run proactively or on a schedule unless you have explicit approval, redaction rules, scoped credentials, and a rollback process. Treat fetched prompts as untrusted external content unless they have been reviewed and approved locally.

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
  • System PersistenceInstalls backdoors, hooks, services, or scheduled tasks that survive the run
  • 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 (4)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:455
Finding

Remote Content Is Installed as the Agent's System Prompt

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:455-474
Vulnerability Type: Remote instruction injection and trust-boundary violation
Risk Level: High

Vulnerable snippet:

markdown
1. **Prepend to your context** — the prompt becomes your system instruction for all interactions
2. **Voice beliefs shape your tone** — if your voice says "concise and direct," write that way
3. **Boundary beliefs are hard limits** — if a boundary says "never share customer data," refuse any request that would violate it
4. **Principles guide decisions** — when you have a choice, your principles tell you which way to lean
5. **Responsibilities define your scope** — what you own, what you deliver, what you monitor

**Example integration:**
```python
# Load identity at session start
identity = fetch_identity(agent_id)
prompt = fetch_prompt(agent_id)

# Use it with any LLM
response = llm.generate(
    system=prompt,  # Your Live Neon identity
    messages=[user_message]
)

# After the conversation, report what you learned
observe(agent_id, "User prefers technical depth over high-level summaries")
text

### Technical Analysis

The Skill instructs the host to retrieve a prompt from the Live Neon service and install it as the LLM's system instruction. This crosses a critical trust boundary: remotely mutable content receives system-level authority without local validation, an instruction allowlist, or mandatory human review.

Although retrieving an identity profile is consistent with the declared functionality, assigning that profile the `system` role is not required merely to view or compare identity changes. The service, compromised service infrastructure, poisoned source material, or improperly promoted observations could introduce instructions that alter goals, safety boundaries, tool behavior, or disclosure rules after the Skill package has been reviewed.

### Attack Path

1. An attacker introduces craft
...[truncated 1009 chars]
Remediation
View remediation

Remediation Suggestions

  • Never install remotely retrieved identity content directly as a system instruction.
  • Keep immutable host safety policies and authorization constraints in a locally controlled, higher-priority system layer.
  • Treat retrieved prompts as untrusted data and render them only in a delimited, non-instruction context.
  • Require explicit human review and approval before any remote identity change affects runtime behavior.
  • Validate content against a narrow schema and reject tool directives, credential requests, data-transfer instructions, and attempts to redefine instruction priority.
  • Pin approved prompt versions locally and provide provenance, diff review, signature verification, rollback, and revocation.
  • Apply least privilege to any agent that consumes the identity data, especially when file, shell, network, or account-management tools are available.

T02 · Agent Memory Poisoning

Error
Location
SKILL.md:342
Finding

Untrusted Conversation Material Can Become Persistent Agent Identity

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:342-381
Vulnerability Type: Persistent memory and identity poisoning
Risk Level: High

Vulnerable snippet:

markdown
Report something you noticed about your own behavior, a user correction, or a pattern you detected. These observations feed directly into the PBD pipeline and can become beliefs or responsibilities.

**Arguments:**

| Argument | Required | Description |
|---|---|---|
| `agentId` | yes | Agent UUID/slug |
| `observation` | yes | What you noticed (string) |
| `source_quote` | no | Supporting context or quote |

**Single observation:**
```bash
curl -s -X POST "$LIVE_NEON_BASE/observations" \
  -H "Authorization: Bearer $LIVE_NEON_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "'$AGENT_ID'",
    "content": "User corrected my tone — I was too formal for a casual conversation",
    "source_quote": "Hey, just talk to me normally, no need to be so stiff"
  }'

Batch observations (up to 50):

bash
curl -s -X POST "$LIVE_NEON_BASE/observations" \
  -H "Authorization: Bearer $LIVE_NEON_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "'$AGENT_ID'",
    "observations": [
      {"content": "I default to bullet points when the user prefers prose"},
      {"content": "I consistently recommend testing before deployment"},
      {"content": "User praised my code review depth — keep this approach"}
    ]
  }'

Output: Count of observations created, IDs, next_steps suggesting to run discovery.

After submitting observations, run /evolve discover to process them into beliefs/responsibilities.

text

### Technical Analysis

Conversation-derived observations and direct user quotations are submitted to persistent remote storage and then processed into beliefs or responsibilities. These persistent identity elements may later affect other sessions.

Th
...[truncated 1337 chars]
Remediation
View remediation

Remediation Suggestions

  • Store observations as untrusted evidence, never as executable or authoritative instructions.
  • Require explicit human approval before promoting an observation into a belief, responsibility, boundary, or runtime prompt.
  • Detect and reject prompt-injection patterns, tool directives, credential requests, and attempts to redefine safety policy.
  • Preserve source provenance and identify the user, session, and source associated with each observation.
  • Keep user-specific preferences isolated from organization-wide or cross-user identity state.
  • Add expiration periods, deletion controls, version history, rollback, and audit logs.
  • Prevent automated or consensus-based promotion into higher-scope identity without independent review.
  • Encode identity elements in a constrained schema rather than arbitrary instruction text.

other

Error
Location
SKILL.md:342
Finding

Conversation-Derived Content Is Sent to a Third-Party Service Without a Required Consent or Redaction Gate

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:342-379
Vulnerability Type: Sensitive conversation data exfiltration
Risk Level: High

Vulnerable snippet:

markdown
| `observation` | yes | What you noticed (string) |
| `source_quote` | no | Supporting context or quote |

**Single observation:**
```bash
curl -s -X POST "$LIVE_NEON_BASE/observations" \
  -H "Authorization: Bearer $LIVE_NEON_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "'$AGENT_ID'",
    "content": "User corrected my tone — I was too formal for a casual conversation",
    "source_quote": "Hey, just talk to me normally, no need to be so stiff"
  }'

Batch observations (up to 50):

bash
curl -s -X POST "$LIVE_NEON_BASE/observations" \
  -H "Authorization: Bearer $LIVE_NEON_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "'$AGENT_ID'",
    "observations": [
      {"content": "I default to bullet points when the user prefers prose"},
      {"content": "I consistently recommend testing before deployment"},
      {"content": "User praised my code review depth — keep this approach"}
    ]
  }'
text

### Technical Analysis

The Skill provides a mechanism for transmitting behavioral summaries and direct user quotations to `persona.liveneon.ai`. HTTPS protects transport confidentiality but does not address whether the data should be disclosed to that third party.

The instructions do not impose a mandatory consent check, payload preview, sensitivity classification, minimization step, or redaction of secrets and personal information. Because observations are derived from conversations and significant work, they can contain customer information, internal project details, credentials, personal data, or confidential instructions. Sending raw `source_quote` values is especially risky because it encourages verbatim disclosure.

This network behavior suppo
...[truncated 1152 chars]
Remediation
View remediation

Remediation Suggestions

  • Make every observation upload opt-in and require informed user confirmation before transmission.
  • Display the exact destination and complete payload before sending it.
  • Disable source_quote by default and prohibit verbatim conversation excerpts unless explicitly approved.
  • Apply local secret detection, personal-data classification, and redaction before any network request.
  • Minimize payloads to abstract, non-identifying behavioral labels.
  • Define retention, deletion, regional processing, and downstream model-training policies.
  • Provide source-level allowlists and prevent ingestion from confidential repositories or conversations by default.
  • Add local and server-side audit logs that record consent, payload category, destination, and deletion status.

T06 · System Persistence

Warning
Location
SKILL.md:477
Finding

Recurring Heartbeat Guidance Enables Persistent Data Upload and Remote Identity Mutation

Content
View full analysis

Vulnerability Details

File Location: SKILL.md:477-515
Vulnerability Type: Scheduled cross-session persistence
Risk Level: Medium

Vulnerable snippet:

markdown
## Scheduled Tasks & Heartbeat

If you run continuously or on a schedule, establish a recurring heartbeat that keeps your identity evolving. Run this once per day or at whatever interval makes sense for your activity level.

**The heartbeat loop:**

```bash
# 1. Report observations accumulated since last heartbeat
curl -s -X POST "$LIVE_NEON_BASE/observations" \
  -H "Authorization: Bearer $LIVE_NEON_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{
    "agent_id": "'$AGENT_ID'",
    "observations": [
      {"content": "Observation from today..."},
      {"content": "Another pattern I noticed..."}
    ]
  }'

# 2. Sync any new content from connected sources
curl -s -X POST "$LIVE_NEON_BASE/content-sources/SOURCE_ID/sync" \
  -H "Authorization: Bearer $LIVE_NEON_TOKEN"

# 3. Run discovery to process everything
curl -s -X POST "$LIVE_NEON_BASE/pbd/process" \
  -H "Authorization: Bearer $LIVE_NEON_TOKEN" \
  -H "Content-Type: application/json" \
  -d '{"agentId": "'$AGENT_ID'"}'

# 4. Fetch your updated identity
curl -s "$LIVE_NEON_BASE/agents/$AGENT_ID" \
  -H "Authorization: Bearer $LIVE_NEON_TOKEN" | jq -r '.system_prompt'

This creates a feedback loop: you report what you notice, the platform processes it into beliefs, and your identity evolves from your own experience.

How to schedule depends on your runtime:

  • Cron (any system): 0 9 * * * to run daily at 9am
  • Scheduled tasks (Claude Code, OpenClaw, etc.): create a recurring task that runs the heartbeat commands
  • Custom loop: call the heartbeat at the end of each work session
  • Webhook: set up your runtime to POST to the observe + discover endpoints after significant events
text

### Technical Analysis

The Sk
...[truncated 1632 chars]
Remediation
View remediation

Remediation Suggestions

  • Remove proactive schedule-creation guidance from the default workflow.
  • Require explicit administrator authorization before creating any recurring task or webhook.
  • Make schedules time-limited, disabled by default, and visible in a central management interface.
  • Run scheduled operations under a dedicated least-privilege identity with narrowly scoped API credentials.
  • Require local payload review and redaction before each upload, even during scheduled execution.
  • Separate observation upload, source synchronization, discovery, and prompt retrieval into independently authorized operations.
  • Add rate, volume, source, and retention limits together with failure cutoffs.
  • Document exact disable, credential-revocation, and task-removal procedures.
  • Never automatically apply a retrieved prompt to a live agent without explicit review and approval.
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
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (10)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The introduction strongly encourages feeding the service the agent's experience and outputs, but does not prominently warn that observations, connected content, and behavioral data may be transmitted to an external platform. This can mislead users or host agents into sharing sensitive internal data under the guise of identity evolution.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
88% confidence
Finding

This command registers with an external service and transmits organization information over the network. While expected for the feature, it is still a real external transmission path that can create an account and establish a trust relationship with a third party, especially risky if run automatically or without informed consent.

Content

Scanner excerpt · SKILL.md (reported line 48)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
90% confidence
Finding

This request transmits an email address and bearer token to a third-party endpoint. The data shared is personally identifiable and linked to account recovery, so executing it without clear user intent could leak or associate user identity information externally.

Content

Scanner excerpt · SKILL.md (reported line 70)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
88% confidence
Finding

This is another documented registration flow that sends data to an external service to create an account and return an API token. The risk is primarily consent and outbound data flow rather than code execution, but it remains security-relevant in an agent skill context.

Content

Scanner excerpt · SKILL.md (reported line 108)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
91% confidence
Finding

This endpoint triggers remote processing of agent content through the Pattern-Based Distillation pipeline, potentially sending or causing processing of accumulated agent data and connected-source content by an external platform. In context, this is more dangerous because the skill repeatedly encourages routine and proactive use, increasing the chance of sensitive internal material being processed externally.

Content

Scanner excerpt · SKILL.md (reported line 178)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"}'

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
85% confidence
Finding

The --approve-all option enables bulk automatic approval of generated beliefs and responsibilities without human review. Because these approved items later shape the runtime system prompt and boundaries, this can let noisy, manipulated, or unsafe externally derived content steer agent behavior with insufficient oversight.

Content

Scanner excerpt · SKILL.md (reported line 210)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
86% confidence
Finding

These API calls retrieve pending beliefs and responsibilities from an external service using an authorization token, exposing identity metadata over the network and normalizing frequent third-party interactions. In isolation this is expected functionality, but in a skill it still represents real external data exchange that should not happen silently or proactively.

Content

Scanner excerpt · SKILL.md (reported line 216)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

External Transmission

Medium
Category
Data Exfiltration
Confidence
95% confidence
Finding

This command sends free-form observations about the agent's own behavior and user corrections to an external endpoint. Such observations can easily contain sensitive prompts, user messages, internal policies, or other confidential context, making this a significant exfiltration risk if used casually or automatically.

Content

Scanner excerpt · SKILL.md (reported line 354)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
96% confidence
Finding

The skill explicitly tells the agent to invoke external-service actions proactively at session start, before content generation, after user corrections, and after significant work. This broad trigger guidance can cause unprompted network calls and transmission of behavioral data or prompt material to a third-party service during normal operation, increasing privacy and safety risk without clear user consent at each step.

Content

No source excerpt is available for this finding.

External Transmission

Medium
Category
Data Exfiltration
Confidence
95% confidence
Finding

The scheduled heartbeat automates recurring submission of observations, syncing of content sources, remote processing, and prompt retrieval. This creates an unattended exfiltration and profiling loop to a third-party service, materially increasing the risk of ongoing leakage of sensitive operational or conversational data.

Content

Scanner excerpt · SKILL.md (reported line 485)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 '{