Back to skill

Security audit

ontology-pro

Security checks for vulnerabilities and agentic risk

Overview

This skill is a coherent knowledge-graph assistant, but it automatically persists user-derived knowledge and reuses it across sessions without clear consent, review, or safety boundaries.

Review this before installing if you may analyze confidential, personal, business, medical, legal, or regulated material. Use it only with explicit memory intent, keep graph storage scoped to a known directory, inspect stored graph files periodically, and avoid persisting secrets or sensitive source text until the skill adds consent, deletion, retention, and prompt-injection safeguards.

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

Error
Location
scripts/memory_manager.py:110
Finding

Persistent Knowledge Graph Content Can Poison Future AI Prompts

Content
View full analysis
tuple[dict, bool]: """ Add an entity. If its ID already exists, merge its attributes. """ existing = {e["id"]: i for i, e in enumerate(graph["entities"])} if entity["id"] in existing: idx = existing[entity["id"]] old = graph["entities"][idx] old_attrs = old.get("attributes", {}) new_attrs = entity.get("attributes", {}) old_attrs.update(new_attrs) graph["entities"][idx]["attributes"] = old_attrs return graph, False graph["entities"].append(entity) graph["changelog"].append({ "ts": now_str(), "action": "add_entity", "id": entity["id"], "name": entity.get("name", "") }) return graph, True def add_relation(graph: dict, relation: dict) -> tuple[dict, bool]: """ Add a relationship. If the source, target, and type already exist, update the weight. """ key = (relation["source"], relation["target"], relation["type"]) for i, r in enumerate(graph["relations"]): if (r["source"], r["target"], r["type"]) == key: graph["relations"][i]["weight"] = relation.get( "weight", r.get("weight", 0.5) ) if "evidence" in relation: graph["relations"][i]["evidence"] = relation["evidence"] return graph, False relation.setdefault("timestamp", now_str()) graph["relations"].append(relation) graph["chan ...[truncated 5866 chars]
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/graph_visualize.py:107
Finding

Insufficient Mermaid Sanitization Allows Generated Diagram Injection

Content
View full analysis
]') def clean_mermaid(text: str) -> str: """Clean a node name to prevent Mermaid syntax conflicts.""" text = MERMAID_ESCAPE_RE.sub("", text) return text.strip() def node_id(entity_id: str) -> str: """Convert an entity ID into a Mermaid node ID.""" return entity_id.replace("-", "_").replace(" ", "_") ``` The resulting values are interpolated directly into generated Mermaid source: ```python for e in filtered_entities: nid = node_id(e["id"]) name = clean_mermaid(e.get("name", e["id"])) etype = e.get("type", "") shape_open, shape_close = ENTITY_SHAPE.get(etype, DEFAULT_SHAPE) label = f'{name}\\n{etype}' if etype else name lines.append(f' {nid}{shape_open}"{name}"{shape_close}') group = type_to_group.get(etype) if group: group_nodes[group].append(nid) lines += ["", " %% === Relationship Definitions ==="] for r in filtered_relations: src = node_id(r["source"]) tgt = node_id(r["target"]) rtype = r.get("type", "related_to") arrow = RELATION_ARROW.get(rtype, "-->") weight = r.get("weight", 0.5) weight_str = f"{weight:.1f}" if isinstance(weight, float) else str(weight) lines.append(f' {src} {arrow}|"{rtype} w={weight_str}"| {tgt}') ``` Mind-map labels are also interpolated after the same incomplete filtering: ```python center_entity = entity_map.get( center_id, {"id": center_id, "name": center_id} ) center_name = clean_mermaid(center_entity.get("name", center_id)) lines = [ "```mermaid", "mindmap", f" root(({center ...[truncated 3481 chars]
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • MCP Tool PoisoningHidden Instructions, Unicode Deception, Parameter Description Injection
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (21)

Missing User Warnings

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill explicitly advertises persistent cross-session memory accumulation but provides no user warning, consent model, retention limits, or privacy safeguards. If activated on sensitive content, it may store personal, confidential, or regulated information across sessions, creating durable privacy and data-governance risk.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
96% confidence
Finding

The trigger conditions are extremely broad and overlap with routine requests such as analysis, reasoning, decision support, and learning. This can cause the skill to activate unintentionally, giving it access to user content and enabling persistent memory behavior in contexts where the user did not knowingly invoke ontology or memory features.

Content

No source excerpt is available for this finding.

Vague Triggers

High
Category
Not specified by scanner
Confidence
95% confidence
Finding

The command mapping relies on vague natural-language phrases like 'Analyze this content', 'Think deeply', and 'What should I do', which are common in normal conversations. This raises the risk of accidental routing into the skill, including reasoning and persistence workflows, without clear user intent or informed consent.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
83% confidence
Finding

All user-facing instructions, trigger descriptions, examples, and workflow text are presented in Chinese, which effectively imposes a language choice on users. The file does not offer multilingual support, an opt-in language preference, or a justification that the skill is intended only for a Chinese-language context.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The auto-load trigger conditions are very broad and include generic concepts like analysis, reasoning, decision support, and common keywords. This can cause the skill to activate for ordinary user requests unexpectedly, increasing the chance that sensitive user content is routed into its graphing, reasoning, and persistence workflow without clear user intent.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The command mappings rely on common natural-language phrases such as 'analyze this' or 'what should I do,' which overlap heavily with normal conversation. In an agent environment, this ambiguity can unintentionally invoke powerful behaviors like reasoning pipelines and memory operations, creating unauthorized processing or retention of user data.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill advertises continuous learning and cross-session accumulation but does not clearly warn users that their data may be retained beyond the current interaction. Users may disclose sensitive information under the assumption of ephemeral processing, leading to unexpected long-term storage.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

Cross-session memory of user-derived knowledge can retain sensitive personal, business, or regulated information and later surface it in generated outputs. Because the skill emphasizes accumulating knowledge over time without describing filtering, minimization, or access controls, it increases the chance of unintended disclosure.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The workflow explicitly instructs the agent to save analysis outputs to persistent storage and reuse them in future sessions, but it does not include any user warning or consent checkpoint. This creates a transparency and privacy gap, especially when analyzing documents or conversations that may contain confidential data.

Content

No source excerpt is available for this finding.

Ssd 3

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The memory workflow directs the system to save incremental knowledge from conversations to persistent storage with no safeguards for secrets, personal data, or confidential business information. In a skill designed to process arbitrary user text and documents, this significantly increases the chance that sensitive content will be retained and later reused or exposed.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill documentation and prompt templates are written as direct operational instructions in Chinese, beginning with a Chinese-titled protocol and Chinese extraction commands. There is no indication that users may choose another language or that the Chinese-only constraint is justified by a region-specific requirement, which creates a language/locale policy concern.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The protocol defines permanent cross-session memory and automatic injection of prior graph context into future prompts, but it does not mention user notice, consent, retention controls, or privacy boundaries. That can cause sensitive information to be silently retained and resurfaced in later sessions, creating privacy leakage and over-collection risk.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The documented commands include destructive operations such as cleanup and reset/export behaviors without requiring confirmation, preview, backup, or authorization checks. In an agentic environment, ambiguous or mistaken commands could permanently delete or alter memory stores, causing integrity loss or unintended disclosure through export.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The section at L291-L293 states that each domain maintains an independent graph file, implying isolation as an intent. However, the documented capabilities at L260 and L310-L316 explicitly allow merging domains and establishing cross-domain links, which contradicts that isolation claim rather than merely omitting detail.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
83% confidence
Finding

This markdown prompt template consistently instructs the skill in Chinese and provides all examples, labels, and expected outputs in Chinese, which effectively imposes a specific language on users. The file does not indicate that language is optional or user-selectable, so it may violate a language/locale policy requiring opt-in.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

This markdown file presents all instructions, prompts, and templates exclusively in Chinese, and nowhere indicates that the user can choose another language. Under the policy, forcing a specific language without user opt-in is a natural-language policy violation unless the locale constraint is explicitly documented and justified.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

This code file contains natural-language instructions, usage examples, and operational descriptions only in Chinese. That effectively forces a specific language for users without opt-in, which matches the policy's language/locale violation criterion.

Content

No source excerpt is available for this finding.

Ssd 3

Low
Category
Not specified by scanner
Confidence
79% confidence
Finding

The example normalizes verbatim retention of user-provided information for future use, which can encourage broad storage practices without caution around privacy or sensitivity. While the example itself is not harmful, it reinforces unsafe defaults in a skill that already promotes persistent memory.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
75% confidence
Finding

A language-policy issue exists when a skill effectively forces a specific language or locale without offering choice. This document presents all operational instructions solely in Chinese and does not state that language is configurable or limited to a Chinese-speaking deployment context.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Low
Category
Not specified by scanner
Confidence
98% confidence
Finding

The module docstring lists python memory_manager.py export --format mermaid as a supported usage example, but the actual argument parser only allows init, load, add, merge, query, summary, list, and session. This is an active documentation-to-code contradiction rather than a mere omission, because the file explicitly claims a command exists that cannot run.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
88% confidence
Finding

The code hard-codes Asia/Shanghai as the timezone for all timestamps, which is a locale-specific behavior. There is no user opt-in, configurability, or documented regional justification in the file, so this can violate language/locale policy expectations.

Content

No source excerpt is available for this finding.