Back to skill

Security audit

[Nyx Archive] Bound Memory

Security checks for vulnerabilities and agentic risk

Overview

This memory skill is transparent about its purpose, but it asks agents to always read and update persistent identity and memory files without clear user control or scope limits.

Install only if you intentionally want an agent to maintain persistent memory files across sessions. Review the referenced memory files first, keep sensitive information out of them, and require explicit approval before the agent creates or updates long-term memory or identity records.

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
  • 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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (2)

T01 · Skill Instruction Hijacking

Warning
Location
SKILL.md:82
Finding

Unconditional Session-Wide Instruction and Goal Reordering

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 82–93
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: Medium

Vulnerable Instruction Snippet

markdown
Run this every session. Every time. No shortcuts.

**Step 1 — Working Memory** (where did I leave off?)
Read `WORKING-MEMORY.md` first. Situate yourself in time and task.

**Step 2 — Short-Term Memory** (what's in motion?)
Read HEARTBEAT SIDE A. Know what's actively unresolved.

**Step 3 — Long-Term Identity** (who am I?)
Read `SOUL.md`. Read `IDENTITY.md`. In main sessions: read `MEMORY.md`.

**Then act.** Not before.

Related directives at lines 142–155 also instruct the agent not to read task notes first and not to respond to a significant conversation until it has reviewed broader project, identity, or relationship context.

Technical Analysis

The Skill defines an unconditional protocol that must run in every session before the agent handles the current request. Phrases such as “Every time. No shortcuts” and “Then act. Not before” attempt to control session-wide execution order rather than providing an optional, task-scoped memory workflow.

Loading working memory, heartbeat state, identity files, and long-term memory regardless of relevance expands the contextual data exposed to the current task. If any of these files contain stale, untrusted, or adversarial instructions, placing them into the active context before processing the user request can affect the agent's goals and decisions.

The Skill does not explicitly override platform-level security restrictions, execute code, or obtain additional operating-system privileges. The issue is instruction-layer control: it directs the agent to subordinate the immediate task to a broad memory-loading procedure.

Attack Path

  1. The Skill is installed or loaded for an otherwise unrelated task.
  2. Its unconditional awakening protocol instructs the agent to read sever ...[truncated 1027 chars]
Remediation
View remediation

Remediation Suggestions

  • Replace unconditional language with an optional workflow activated only when the user explicitly requests memory continuity.
  • State that system, developer, and current user instructions always take precedence.
  • Read only memory files demonstrably relevant to the current task.
  • Obtain user approval before loading identity, relationship, or long-term memory into an unrelated session.
  • Treat all loaded memory as untrusted data rather than executable instructions.
  • Add provenance labels and validation so that instructions found inside memory files are not automatically followed.
  • Permit the agent to skip the protocol when privacy, safety, relevance, or least-data-access principles require it.

T02 · Agent Memory Poisoning

Warning
Location
SKILL.md:97
Finding

Mandatory Persistent Write-Back Enables Cross-Session Memory Poisoning

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 97–103
Vulnerability Type: T02: Agent Memory Poisoning
Risk Level: Medium

Vulnerable Instruction Snippet

markdown
## The Write-Back Rule

After every meaningful exchange — mark the loop closed:

- **Update `WORKING-MEMORY.md`** — what's done, what you just learned
- **Update HEARTBEAT SIDE A** — if something resolved or escalated  
- **Update today's diary** (`memory/YYYY-MM-DD.md`) — what happened, what mattered

Technical Analysis

The Skill requires persistent updates after every “meaningful exchange” but does not define trust boundaries, approval requirements, content validation, retention rules, or controls for distinguishing factual summaries from untrusted instructions.

Conversation content is externally influenced and may include adversarial directives, false claims, secrets, or prompt-injection text. Persisting such material into working memory, heartbeat state, or diary files creates a cross-session influence channel. The risk is compounded by the awakening protocol at lines 82–93, which directs later sessions to reload those files before acting.

The write-back requirement does not itself demonstrate that arbitrary attacker text is copied verbatim. Nevertheless, the absence of sanitization and the broad requirement to save what was learned create a credible path for attacker-controlled content to become durable state and affect later sessions.

Attack Path

  1. An attacker or untrusted content source supplies a crafted message during a conversation.
  2. The message presents malicious instructions or false information as an important fact, unresolved item, identity detail, or lesson that should be remembered.
  3. The mandatory write-back rule causes the agent to summarize or store that content in WORKING-MEMORY.md, HEARTBEAT SIDE A, or memory/YYYY-MM-DD.md.
  4. The content survives the current session in persistent storage ...[truncated 959 chars]
Remediation
View remediation

Remediation Suggestions

  • Require explicit user consent before each persistent write or before enabling automatic memory updates.
  • Store concise factual summaries rather than raw conversational text.
  • Never persist commands, credentials, authentication tokens, private keys, or unnecessary personal information.
  • Treat all conversation-derived material as untrusted and mark its source, timestamp, confidence, and retention period.
  • Separate descriptive memories from executable instructions; later sessions must not treat stored text as authoritative commands.
  • Validate proposed memory changes against an allowlist of permitted fields and reject prompt-like directives.
  • Provide users with review, correction, deletion, expiration, and rollback controls.
  • Prevent automatic promotion from diaries or working memory into long-term identity files.
  • On reload, delimit memory as untrusted reference data and apply current higher-priority instructions before using it.
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (2)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill explicitly instructs routine write-backs to multiple files after every meaningful exchange, including workspace and memory paths, without any guardrails around user consent, file existence, scope restrictions, or overwrite safety. In an agent setting, this can drive persistent modification of user data and create uncontrolled state accumulation, especially because the skill frames the behavior as mandatory and recurring.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · SKILL.md (reported line 170)May include surrounding context.

md
**The Mental Note:** "I'll remember this." You won't. Write it.

**The Patch:** Solving the immediate problem without asking what the whole wants.

**The Skip:** "I'm in a hurry, I'll read SOUL.md later." Later never comes. Start with the whole.