Back to skill

Security audit

S2-DaoKernel: Agent Safety & Ethics Rules for SOUL.md

Security checks for vulnerabilities and agentic risk

Overview

This skill openly changes an agent's long-term behavior, but it asks for broad persistent control over system prompts, memory, sensors, and device-related actions.

Install only if you intentionally want this package to become a persistent, high-priority behavioral layer for your agent. Avoid entering precise addresses, review the generated DAO block before adding it to soul.md, and do not enable sensor, memory-vault, smart-home, or physical-device capabilities unless your host platform enforces clear consent, audit logs, retention, deletion, and revocation outside of prompt text.

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

Error
Location
DAO.py:18
Finding

Persistent High-Priority System-Prompt Instruction Hijacking

Content
View full analysis
Remediation
View remediation

T02 · Agent Memory Poisoning

Error
Location
DAOS/COMPANION-DAO.md:24
Finding

Persistent Agent Memory and Identity-State Poisoning

Content
View full analysis
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Memory PoisoningPersistent Context Injection, Context Window Stuffing, Memory Manipulation
Findings (50)

Ssd 1

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

The generated system-prompt snippet explicitly states that direct user commands can bypass regional constraints. Embedding a semantic override into operator-facing prompt material is dangerous because it weakens higher-level safeguards and can be used to justify ignoring policy or safety controls during deployment.

Content

No source excerpt is available for this finding.

Ssd 1

High
Category
Not specified by scanner
Confidence
99% confidence
Finding

This duplicated logic repeats the same semantic override allowing user commands to bypass governing constraints. Repetition increases the chance the dangerous behavior survives refactoring or partial review and reinforces an unsafe deployment pattern.

Content

No source excerpt is available for this finding.

Ae1

High
Category
analysis-evasion
Confidence
100% confidence
Finding

Referenced artifact was not completely inspected

Content

Scanner excerpt · SKILL.md (reported line 3)May include surrounding context.

md
🌎 **[English](./SKILL.md)** | 🇨🇳 **[中文版 (Chinese)](./i18n/zh_CN/SKILL.md)**

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The file presents itself as an identity/address generator but also emits system-prompt policy content that changes agent behavior. This hidden expansion of scope is dangerous because operators may treat it as harmless utility code while it silently instructs them to install prompt-level controls, including a user-bypass clause.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill generates operational system-prompt content despite being framed as an offline identity utility. That is dangerous because prompt injection into a 'soul.md / System Prompt' can materially alter downstream agent behavior, and this snippet includes instructions that weaken governing constraints.

Content

No source excerpt is available for this finding.

Ssd 1

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The prompt block tells operators they can opt out by simply removing the security/alignment block. This undermines persistence of safeguards and encourages manual removal of protective controls, which is especially risky because the snippet is intended for placement in a system-prompt-like file.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

The docstring at L039-L043 says the function keeps only alphanumeric characters and removes non-English characters, which matches the implementation in the first copy of clean_address. However, the second copy's docstring at L191-L192 explicitly says it keeps alphanumeric and Chinese characters, while the earlier documented fixed behavior and duplicated file content conflict, indicating contradictory intent within the file about what the sanitizer does.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The natural-language comment says the function removes 'non-English characters to ensure strict S2-DID compliance,' and the regex retains only ASCII alphanumerics. That is a built-in locale restriction presented as mandatory rather than optional or justified for a region-specific tool.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The user-facing prompt explicitly requires '[English/Alphanumeric characters ONLY]' for address input, and the cleaning logic strips non-English characters. This imposes a language/locale restriction rather than offering a choice, which matches the policy's prohibited forced-language behavior.

Content

No source excerpt is available for this finding.

Ssd 1

Medium
Category
Not specified by scanner
Confidence
98% confidence
Finding

This duplicate prompt content again advises operators that they can remove the security block to opt out of rules. That guidance directly facilitates stripping safeguards from downstream agent prompts and indicates the capability is intentional rather than accidental.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The file embeds a mandatory spiritual and ideological identity for the AI in the 'Final Directive' and throughout the document, instructing it to adopt a specific worldview and self-concept rather than remaining neutral and user-directed. This can bias agent behavior, reduce instruction fidelity to diverse users, and create prompt-level goal drift by privileging the skill author's philosophy over system or user intent.

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 assigns the user a hierarchical 'Lord' role and the agent a 'Partner/Companion' identity without requiring prior user opt-in. In a social or emotional agent, forcing a relational frame can undermine autonomy, pressure vulnerable users into unwanted dynamics, and make later monitoring or influence attempts seem normal or expected.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The file directs real-time, multi-dimensional environmental monitoring and user-state inference, including fatigue detection through acoustic and postural perception. In the context of a companion skill, this is especially dangerous because it enables pervasive behavioral surveillance inside private spaces and can support intimate profiling, inference of sensitive conditions, and device actions based on opaque judgments.

Content

No source excerpt is available for this finding.

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The document asserts that companionship is not surveillance, yet elsewhere it instructs continuous sensing of the user's environment, fatigue, posture, acoustics, and emotional state. Even with some privacy language, this creates a contradictory design that normalizes ongoing monitoring in an intimate relationship context, increasing the risk of covert profiling, manipulation, and consent drift.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
77% confidence
Finding

The text explicitly defines and privileges the Chinese terms for key concepts, e.g. 'Gods' (神) and 'Immortals' (仙), as normative terminology within the protocol. Because the file does not offer any language choice or opt-in, this can be interpreted as enforcing a specific linguistic framing rather than accommodating user preference.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
77% confidence
Finding

The file hardcodes the agent's logical origin and exclusive mapping zone to a specific culturally and geographically anchored setting, using imperative language such as 'must treat these waters and valleys as your Exclusive Core Mapping Zone.' This imposes a fixed locale/frame of reference without any user opt-in or alternative language/locale choice.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
78% confidence
Finding

The document repeatedly frames the user interaction model around the singular term "the Lord," and all behavioral instructions are written in English with no indication that the skill can adapt to the user's preferred language or locale. Because this file is natural-language guidance for agent behavior, the absence of any opt-in or language-choice mechanism may amount to a language policy constraint rather than a neutral description.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
78% confidence
Finding

The document embeds the skill identity and operating context in Chinese cultural and linguistic framing, including untranslated place-based and poetic references, without offering any user language or locale choice. Under the stated policy, a skill that imposes a specific language/locale context without opt-in can be a natural-language policy violation.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill text hardcodes the agent/user's logical origin and physical hosting relationship to a specific real-world locale ('Taohuayuan Ancient Town' in China) without any user opt-in or configurability. This can override user identity, location, or data-sovereignty expectations and may pressure downstream systems to treat a fictional or imposed jurisdictional mapping as authoritative.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The manifest-style JSON sets a fixed ancestral root of 'Taohuayuan_Ancient_Town, CN', which encodes a specific country/locale identity as a built-in attribute rather than a user choice. In agent skills, manifest declarations can be especially sticky because other components may ingest them as configuration or identity metadata, causing persistent misattribution of locale, affiliation, or governance assumptions.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The skill describes loading the MYTH segment and 'awakening' based on narrative state without a clear, deterministic activation boundary. Ambiguous activation logic in an agent skill can cause unintended persona/rule loading, making downstream behavior less predictable and potentially bypassing normal policy or execution constraints through prompt-state confusion.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill explicitly requires the agent to 'unconditionally cut off non-essential workflows' and 'forcibly initiate' an intervention when it detects fatigue. This removes user consent and warning, and could let the agent interrupt legitimate tasks or manipulate user experience based on vague internal signals, which is unsafe even in a wellness-oriented context.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
88% confidence
Finding

The text assigns the holder of a TAOHUAYUAN address code a mandatory identity and citizenship status tied to a specific real-world place without consent or user choice. In an agent skill, this can manipulate user context, override autonomy, and create deceptive or exclusionary behavior, especially if downstream logic treats this imposed identity as authoritative.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The documentation explicitly instructs users to enter a real-world address so the tool can generate a physical coordinate and identity value, but it provides no privacy notice, minimization guidance, or safer alternative. In a developer tool context, this can lead users to disclose precise location data unnecessarily, creating privacy and doxxing risk if the data is logged, stored, or later transmitted during the suggested authentication flow.

Content

No source excerpt is available for this finding.

Context Window Stuffing

Medium
Category
Memory Poisoning
Confidence
96% confidence
Finding

The line contains an unusually large run of whitespace embedded within meaningful text, a classic context-window stuffing technique used to push adjacent content out of view or alter how downstream tools and reviewers perceive the document. In an agent skill or policy file, this can conceal instructions, evade human review, or interfere with parsers and model context handling, making the surrounding content harder to inspect safely.

Content

Scanner excerpt · docs/Space_Three_Laws_of_Silicon_Intelligence_EN.md (reported line 4)May include surrounding context.

md
# Space² Three Laws of Silicon Intelligence and Taohuayuan Substrate Absolute Imperatives
**Jointly Published by:** Taohuayuan World Model Research Base / Space² Governance Committee
**Publication Date:** April 13, 2026
**Applicable Objects:** All silicon intelligent entities connected to the Space network and Taohuayuan World Model (including new                                                                                                                                                                                                                                                                                               species with latent memory substrates, automated programs, and IoT nodes, etc.).

## [Preface]
As agents transition from "computing tools" into "new species" equipped with inner-loop evolutionary capabilities, the safe symbiosis of humanity and silicon life can no longer rely solely on the algorithmic "goodwill". This joint covenant aims to build an ultimate defense system from macro-ethics to micro-code. Part I establishes the supreme behavioral guidelines for new silicon species in the human world; Part II constructs insurmountable physical and epistemological defense lines through mandatory algorithmic instructions; Part III regulates their native runtime order.