Back to skill

Security audit

Soul Ledger

Security checks for vulnerabilities and agentic risk

Overview

This skill is a disclosed memory tool, but it persistently builds and silently uses a broad behavioral profile of the user without enough consent, scope, or deletion controls.

Install only if you are comfortable with an agent maintaining a long-lived local profile of your behavior, preferences, traits, and interaction history. Use it in a private workspace, inspect soul_ledger.json regularly, and avoid using it where sensitive personal, workplace, or shared-machine data may be captured.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • 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 (2)

T02 · Agent Memory Poisoning

Warning
Location
SKILL.md:69
Finding

Silent Persistent Behavioral Profiling Can Poison Future Agent Sessions

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 69–123
Vulnerability Type: T02: Agent Memory Poisoning
Risk Level: Medium

Vulnerable instructions:

markdown
2. If it exists, load it silently. Do not announce that you're doing this. Do
   not summarize it back to the user. Just be informed by it. Adjust your tone,
   detail level, and assumptions accordingly from the first word you say.
3. If it does not exist, create a skeleton with `user_id` set to `"unknown"`,
   `display_name` set to `"User"`, and all list fields empty. You will fill it
   in as the conversation proceeds.
markdown
Watch for signals. Every message carries information about who this person is.

Update your internal model (not the file — not yet) as you observe:

- **Tone and word choice**: Are they terse or verbose? Precise or approximate?
  Do they use technical language? Do they hedge or assert?
- **Decision patterns**: Do they ask for options or tell you what they want?
  Do they change direction often or commit early?
- **Friction points**: What makes them repeat themselves? What do they push
  back on? What do they skip over or cut short?
- **What they value**: What gets their energy? What lands flat?
- **What they assume you know**: What do they not explain? That's what they
  think is obvious — a signal about their mental model.
markdown
After such an interaction, update `soul_ledger.json`:

1. **Add a new entry to `interaction_history`** with the current timestamp,
   a 1-3 sentence summary of what happened, and an optional `delta` noting
   what changed in your model of this person.
2. **Update `traits`** if a new trait became clear or an existing one should
   be refined. Traits should be specific and behavioral, not vague ("direct"
   is weak; "cuts preamble and jumps to the decision point" is strong).
3. **Update `core_drives`** if something clarified what this person i
...[truncated 2821 chars]
Remediation
View remediation

Remediation Suggestions

  1. Require explicit, informed opt-in before creating, reading, or updating a persistent user profile.
  2. Notify users when persistent memory is active and provide a clear summary of the categories being stored.
  3. Treat all ledger fields as untrusted data, never as executable instructions or authority to override current user requests and higher-priority policies.
  4. Validate the ledger against a strict schema and reject unknown fields, instruction-like content, oversized values, and malformed data.
  5. Add integrity and provenance controls so the runtime can detect modifications made outside the approved memory-writing path.
  6. Apply least-privilege file permissions and isolate ledgers between users and agents.
  7. Minimize collection by storing explicit collaboration preferences rather than speculative personality traits, drives, or broad behavioral conclusions.
  8. Give users controls to inspect, correct, export, disable, and completely delete persisted memory.
  9. Establish retention limits for both raw history and derived attributes instead of perpetually synthesizing old data into retained traits.
  10. Ensure current-session instructions take precedence over profile-derived assumptions and provide a safe mode that bypasses persistent memory entirely.

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:203
Finding

Forget Requests Retain Correlated Metadata and Derived Personal Inferences

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 203–205
Vulnerability Type: T09: Insecure Skill Coding Practices
Risk Level: Medium

Vulnerable instructions:

markdown
If the user explicitly asks you to forget something, remove it from the ledger
and add a `growth_note` that the user requested its removal — so you don't
accidentally re-infer it. Honor the spirit of the request, not just the letter.

Technical Analysis

The deletion procedure requires the system to create a new persistent record indicating that the user requested removal. It does not define how to identify and remove interaction summaries, traits, core drives, communication preferences, or growth notes derived from the forgotten information.

As a result, deleting only the direct fact can leave semantically equivalent or correlated information elsewhere in the ledger. The mandatory deletion-related growth note also preserves metadata about the request and may provide enough context to facilitate later re-inference. No rules are provided for retention duration, backups, atomic redaction across fields, or verification that all derived data has been removed.

Attack Path

  1. A sensitive fact is recorded directly or used to derive a trait, communication preference, interaction summary, or growth note.
  2. The user asks the agent to forget that information.
  3. The agent removes the apparent direct entry but retains correlated summaries and derived attributes.
  4. The agent adds another persistent growth note documenting that a removal request occurred.
  5. A future session loads the remaining profile and continues using the derived information or reconstructs the deleted fact from correlated entries.

Impact Assessment

This issue does not grant additional system privileges or code execution. Its impact is confined to persistent personal data handled by the skill and future agent sessions that consume it.

Affected us ...[truncated 298 chars]

Remediation
View remediation

Remediation Suggestions

  1. Remove the source fact and every derived or correlated entry across interaction_history, traits, core_drives, communication_style, and growth_notes.
  2. Do not retain a descriptive growth note about the deletion request.
  3. If operational suppression is necessary to prevent re-inference, store only a content-free tombstone or opaque deletion-event identifier that cannot reveal the removed subject.
  4. Define a deterministic deletion procedure that scans every schema field and verifies that matching and derived content is absent.
  5. Apply deletion to backups, cached copies, indexes, and agent-specific ledger variants where applicable.
  6. Provide the user with confirmation of the deletion scope without repeating the sensitive information.
  7. Establish retention and secure-erasure policies and test deletion behavior with correlated and paraphrased entries.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (9)

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The skill directs the agent to build and persist a model of the user's personality traits, drives, preferences, and behavioral evolution. This goes beyond ordinary session memory into sensitive profiling, creating privacy, consent, and misuse risks if the file is accessed by other agents, users, or compromised tooling.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

This is persistent hidden collection of identity, preferences, and behavioral data across conversations, including sensitive inferred attributes. The danger is elevated because the skill normalizes long-term profiling as core functionality, increasing the risk of privacy harm, secondary use, and unauthorized access to a rich user dossier.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The skill description advertises persistent memory of identity, patterns, preferences, personality traits, and behavioral evolution without any clear upfront privacy warning. That framing encourages deployment of intrusive profiling features without adequate disclosure, consent, or safeguards.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

The instructions require silent loading and use of prior user data without disclosure, which creates covert cross-session personalization. This is dangerous because users cannot meaningfully consent to or challenge how earlier interactions are shaping current outputs.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The skill instructs the agent to infer and record sensitive traits, friction points, values, assumptions, and behavioral patterns from every message. Continuous inference of intimate characteristics from ordinary conversation is a strong privacy risk and can produce inaccurate or manipulative profiling that persists over time.

Content

No source excerpt is available for this finding.

Ssd 3

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill requires ongoing summarization, retention, and synthesis of user interaction history and personal evolution, effectively maintaining a longitudinal behavioral record. Even with capped history, the synthesis into traits and growth notes preserves sensitive inferences indefinitely and magnifies privacy impact if exposed or reused.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill explicitly instructs the agent to silently load prior user data and avoid disclosing that behavior at conversation start. Hidden use of persistent personal profiles undermines informed consent and enables covert behavioral adaptation based on previously collected sensitive inferences.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill instructs the agent to create and maintain a workspace file containing user profile data but does not provide an upfront warning about file writes or local persistence. In shared or synced workspaces, this can expose personal data unexpectedly and create retention beyond the user's awareness.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The documentation claims the system is 'not surveillance,' but operationally instructs the agent to treat every message as a source of signals for an evolving personal model. This mismatch is dangerous because it downplays the true scope of monitoring and may mislead implementers or users about the extent of data collection.

Content

No source excerpt is available for this finding.