Back to skill

Security audit

Memory Stack Core

Security checks for vulnerabilities and agentic risk

Overview

This memory skill is purpose-aligned but broadly and automatically stores conversation details and full exchanges for later reuse, creating privacy and persistent prompt-injection risks.

Install only if you are comfortable with broad local persistence of conversation content. Use it in workspaces where stored prompts, paths, URLs, decisions, and assistant outputs are acceptable to keep on disk; avoid it for secrets or sensitive client data unless you add opt-in capture, redaction, review, and deletion controls.

Vulnerability Patterns
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
  • 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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
Findings (1)

T02 · Agent Memory Poisoning

Error
Location
scripts/run.py:74
Finding

Persistent Prompt Injection Through Untrusted Memory Capture and Recovery

Content
View full analysis
max_bytes: # Rotate: keep last max_size/2 all_text = BUFFER_FILE.read_text() half = max_bytes // 2 BUFFER_FILE.write_text(all_text[-half:]) except: pass ``` `scripts/run.py:113-131` automatically extracts attacker-controlled text and stores it in the WAL. It also stores the first 100 characters of the original message as context: ```python def auto_capture_wal_from_message(message: str): """Scan message for specifics and write to WAL automatically.""" if not load_config()["wal"].get("auto_capture", True): return patterns = { "path": r"(/[-\w./]+|~[-\w./]+|[a-zA-Z]:\\[-\w.\]+)", "url": r"https?://[^\s]+", "decision": r"(?:decide|choose|go with|let's|we will)\s+([^\n.,;]+)", "preference": r"(?:prefer|like|don't like|hate|rather)\s+([^\n.,;]+)", "value": r"([\d]+(?:\.\d+)?)(?:\s*(?:seconds?|minutes?|hours?|days?|GB|MB|KB))?", "correction": r"(?:actually|correction|oops|sorry,\s*)\s*([^\n.,;]+)" } found = False for cat, pat in patterns.items(): for m in re.finditer(pat, message, re.IGNORECA ...[truncated 3755 chars]
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 Least PrivilegeUnderdeclared Capability, Wildcard Permission, Missing Permission Declaration
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
Findings (8)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill promotes automatic persistence and recovery features but does not prominently warn users that it stores full exchanges and specific user-provided details in workspace files. That omission undermines informed consent and can lead users to disclose sensitive information that is later retained and resurfaced across sessions.

Content

No source excerpt is available for this finding.

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
88% confidence
Finding

The skill performs persistent reads and writes to workspace memory files, but the top-level skill metadata does not declare an overall tool scope such as permissions or allowed-tools. That weakens transparency and reviewability, making it easier to install a skill that silently persists user data beyond what a user may reasonably expect.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The WAL auto-capture rule is so broad that it would trigger on routine user messages containing names, numbers, preferences, paths, URLs, or corrections. This creates indiscriminate persistence of ordinary conversation details, increasing the chance that secrets, personal data, or sensitive project information are written to disk without meaningful user intent or consent.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The WAL guidance directs automatic recording of specifics from human messages, including names, IDs, URLs, paths, values, and surrounding context snippets. Those fields are exactly the kinds of data that often contain credentials, personal data, internal infrastructure details, or sensitive business context, so automatic capture materially increases confidentiality risk.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The instructions explicitly tell the agent to persist full user/assistant exchanges and later recover them after compaction or restart. This creates a natural-language data retention channel that can preserve secrets, credentials, personal data, or proprietary content outside the transient context window and reintroduce them later.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

buffer_write appends full content by role directly into working-buffer.md, which can include unrestricted user-provided text and agent outputs. In a memory-resilience skill, this broad retention is especially risky because the stated purpose is to survive compaction and restarts, meaning sensitive content is intentionally kept longer and made easier to recover from disk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill automatically persists message-derived content to local files (WAL and buffer) without any built-in consent, disclosure, or data-minimization boundary. In an agent memory component, this creates a real privacy and secret-retention risk because arbitrary user prompts may contain credentials, personal data, internal paths, or sensitive decisions that are then silently written to disk.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The auto_capture_wal_from_message function scans natural-language messages for paths, URLs, decisions, preferences, values, and corrections, then stores matches automatically. This is dangerous because it turns ordinary conversation into persistent structured telemetry, increasing the chance that sensitive user data or operational details are retained and later exposed through local access, logs, backups, or downstream tooling.

Content

No source excerpt is available for this finding.