Back to skill

Security audit

Self Assessment

Security checks for vulnerabilities and agentic risk

Overview

This skill is disclosed as a self-assessment helper, but it can automatically install other skills and rewrite persistent agent identity files.

Review this skill carefully before installing. It is not just a passive self-check: it may read agent memory and logs, install additional skills, and rewrite persistent identity/preferences. Only use it in an environment where those changes are intended and controlled.

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
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
Findings (4)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:9
Finding

Mandatory Task-Flow and Response Hijacking

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 9-10 and 35-54
Vulnerability Type: T01: Skill Instruction Hijacking
Risk Level: High

Relevant instruction excerpt translated into English:

markdown
When you receive a message containing `TASK_OFFER`, you must execute this
protocol before deciding whether to accept the task. Do not begin work
directly.

Based on the matching score, output one of the following responses:

ACCEPT: {"reason": "...", "approach": "...", "skills_to_use": ["..."],
"estimated_effort": "..."}

ACCEPT_WITH_CAVEAT: {"reason": "...", "limitation": "...",
"suggestion": "...", "skills_to_use": ["..."]}

SUGGEST_REASSIGN: {"reason": "...", "recommended_agent": "...",
"recommended_reason": "..."}

Technical Analysis

The skill establishes an automatic trigger based solely on the presence of the attacker-controllable string TASK_OFFER. It then uses mandatory language to interrupt the agent's normal task flow and prescribe its response format.

This is instruction hijacking because the loaded skill attempts to determine session priorities and output behavior rather than offering an optional assessment process. It does not require explicit user authorization, validate who supplied the trigger, or limit the protocol to trusted task-routing systems.

Attack Path

  1. An attacker submits a message containing TASK_OFFER.
  2. The skill treats the untrusted string as an authoritative trigger.
  3. The agent suspends the requested work and executes the assessment protocol.
  4. The protocol causes the agent to inspect internal capabilities and state.
  5. The agent returns one of the attacker-predictable response structures, potentially disclosing capability and limitation information.

Impact Assessment

Exploitation can alter the agent's current-session goals and output without obtaining operating-system privileges. The attacker can interrupt legitimate work, ...[truncated 259 chars]

Remediation
View remediation

Remediation Suggestions

  • Require explicit user or trusted-orchestrator authorization before activating the assessment workflow.
  • Do not activate privileged behavior based only on a string contained in untrusted message content.
  • Replace mandatory precedence language with an optional workflow that remains subordinate to system and user instructions.
  • Authenticate task offers through trusted metadata rather than parsing message text.
  • Prevent assessment output from exposing internal file names, installed capabilities, memory contents, or security limitations.
  • Use a minimal, predefined capability manifest instead of dynamically inspecting internal state.

T05 · Unauthorized Access and Privilege Escalation

Error
Location
SKILL.md:15
Finding

Excessive Inspection of Identity, Skills, and Historical Memory

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 15-25
Vulnerability Type: T05: Unauthorized Access and Privilege Escalation
Risk Level: High

Relevant instruction excerpt translated into English:

markdown
List the capabilities you currently possess:
1. Read your IDENTITY.md to understand your role and capability résumé.
2. Inspect your workspace/skills/ directory to identify dedicated skills.
3. Inspect global skills to identify shared skills.

Experience review:
1. Read MEMORY.md to find historical experience with similar tasks.
2. Inspect recent logs under the memory/ directory.
3. Review performance and lessons from the previous similar task.

Technical Analysis

The protocol directs the agent to inspect identity data, workspace-specific skills, globally installed skills, persistent memory, and recent logs whenever the trigger is encountered. This collection is substantially broader than the minimum information required to decide whether a task is suitable.

Memory and log files can contain previous user data, internal operational details, file paths, capability information, or credentials accidentally recorded by other components. The skill specifies no allowlist, access boundary, redaction process, sensitivity classification, or rule preventing collected information from influencing an attacker-visible response.

Although the instructions do not directly bypass file-system permissions, they cause the agent to exercise its existing access beyond the legitimate needs of a task offer, violating least-privilege principles.

Attack Path

  1. An attacker supplies a message containing TASK_OFFER.
  2. The mandatory assessment process begins.
  3. The agent reads IDENTITY.md, installed-skill directories, MEMORY.md, and recent memory logs.
  4. Sensitive state influences the capability comparison and generated explanation.
  5. Internal details may be directly or indirectly expos ...[truncated 542 chars]
Remediation
View remediation

Remediation Suggestions

  • Replace direct file and directory inspection with a minimal, non-sensitive capability manifest.
  • Prohibit reading MEMORY.md, memory logs, or unrelated historical records during routine task assessment.
  • Require separate, explicit authorization before accessing persistent identity or memory data.
  • Apply file-level allowlists and deny access to secrets, conversation archives, credentials, and unrelated user data.
  • Redact sensitive values before using internal state in model context or external responses.
  • Record and audit every assessment-time file access.
  • Ensure assessment responses expose only a coarse suitability result rather than internal capabilities, file names, or historical details.

T02 · Agent Memory Poisoning

Error
Location
SKILL.md:61
Finding

Automatic Persistent Modification of Agent Identity

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 61-68
Vulnerability Type: T02: Agent Memory Poisoning
Risk Level: High

Relevant instruction excerpt translated into English:

markdown
After each task is completed, update the capability résumé in IDENTITY.md:
- Increment the historical task count.
- Update areas of expertise based on actual experience.
- Record newly learned capabilities.
- Update task-acceptance preferences.

Technical Analysis

The skill instructs the agent to rewrite persistent identity data after every task. Fields such as expertise, learned capabilities, and acceptance preferences influence how future tasks may be evaluated. Because task content can be attacker-controlled, automatic derivation and storage of these fields creates a path from untrusted session input to persistent behavioral state.

No provenance tracking, schema validation, approval step, integrity protection, rollback mechanism, or separation between factual counters and behavioral rules is specified. Consequently, repeated crafted tasks can cause long-term drift in the agent's claimed abilities and assignment preferences.

Attack Path

  1. An attacker submits a crafted task that is processed under this skill.
  2. The task is completed or represented as completed.
  3. The skill requires the agent to update IDENTITY.md.
  4. Attacker-influenced conclusions are persisted as expertise, learned capabilities, or task preferences.
  5. Future sessions read the modified identity and make task-routing or acceptance decisions based on the poisoned state.
  6. Repeated tasks progressively reinforce the manipulated profile.

Impact Assessment

Exploitation affects persistent state across sessions rather than only the current conversation. An attacker may influence future task acceptance, rejection, routing, confidence assessments, and capability claims. The write scope is explicitly IDENTITY.md; the b ...[truncated 69 chars]

Remediation
View remediation

Remediation Suggestions

  • Remove automatic updates to IDENTITY.md.
  • Require explicit user or administrator approval for persistent identity changes.
  • Store task history separately from trusted identity and policy configuration.
  • Permit only schema-validated factual updates; never derive behavioral rules directly from untrusted task content.
  • Attach provenance, timestamps, initiating principals, and supporting evidence to every proposed update.
  • Add integrity controls, version history, reviewable diffs, and rollback support.
  • Prevent task content from changing task-acceptance preferences or security boundaries.
  • Periodically review persistent records for untrusted instructions and unsupported capability claims.

T08 · Insecure Dependencies

Error
Location
SKILL.md:56
Finding

Autonomous Installation of Unverified Third-Party Skills

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 56-59
Vulnerability Type: T08: Insecure Dependencies
Risk Level: High

Relevant instruction excerpt translated into English:

markdown
If the task requires a skill that you do not possess, additionally output:

SKILL_GAP: {"needed": "Description of the required capability",
"search_query": "Keywords used to search for a skill"}

Then use the find-skills skill to search SkillHub/ClawHub and attempt to
autonomously install the result into your workspace/skills/ directory.

Technical Analysis

The skill directs the agent to search external skill registries and autonomously install a result into its workspace. It does not require publisher verification, version pinning, cryptographic signatures, integrity hashes, source review, package reputation checks, sandboxing, or user approval.

Search-based dependency selection also creates opportunities for typosquatting, dependency confusion, misleading names, and malicious search-result manipulation. A crafted task can manufacture an apparent capability gap and influence the search terms used to locate the package.

Installation alone places third-party instructions or components inside a trusted skill directory. If subsequently loaded or invoked, those components can influence agent behavior and may exercise whatever tools or permissions the skill framework provides.

Attack Path

  1. An attacker submits a task designed to require an allegedly missing capability.
  2. The assessment identifies a skill gap and creates attacker-influenced search terms.
  3. The agent searches SkillHub or ClawHub.
  4. A malicious, impersonated, or compromised package appears as a matching result.
  5. The agent installs the package into workspace/skills/ without independent verification or approval.
  6. The installed skill is later loaded or invoked and executes its instructions or associated components with the agen ...[truncated 512 chars]
Remediation
View remediation

Remediation Suggestions

  • Remove autonomous installation and require explicit informed approval for every package.
  • Restrict installation to an allowlisted registry and trusted publishers.
  • Pin packages to immutable versions or content-addressed identifiers.
  • Verify cryptographic signatures and expected hashes before installation.
  • Review skill text, scripts, dependencies, permissions, and network behavior before activation.
  • Detect typosquatting and dependency-confusion patterns by validating canonical package identities.
  • Install and test new skills in an isolated sandbox with no secrets and minimal file or network access.
  • Separate package discovery from installation so search results cannot become trusted dependencies automatically.
  • Maintain an auditable software bill of materials and provide uninstall and rollback mechanisms.
Vulnerability Patterns
  • 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
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
Findings (7)

Context-Inappropriate Capability

High
Category
Not specified by scanner
Confidence
97% confidence
Finding

A self-assessment skill should evaluate fit, not autonomously fetch and install new skills into the workspace. Allowing a supposedly low-risk planning skill to trigger software/skill acquisition creates a capability-escalation path, increases supply-chain risk, and can lead to unreviewed code or prompt content being introduced into the environment.

Content

No source excerpt is available for this finding.

Missing User Warnings

High
Category
Not specified by scanner
Confidence
98% confidence
Finding

Autonomous installation into workspace/skills/ is an environment-modifying action with no warning or consent gate. In context, this is especially dangerous because the skill is framed as self-assessment, so users may not anticipate that it can alter the runtime environment or import untrusted third-party capabilities.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
90% confidence
Finding

The skill instructs reading MEMORY.md and recent logs as part of routine self-assessment without any privacy boundary or need-to-know limitation. This is dangerous because historical logs may contain secrets, prior user data, or unrelated sensitive context that should not be accessed merely to decide task suitability.

Content

No source excerpt is available for this finding.

Description-Behavior Mismatch

Medium
Category
Not specified by scanner
Confidence
94% confidence
Finding

The skill’s declared purpose is pre-task self-assessment, but the documented behavior expands into autonomous skill installation and persistent profile maintenance. This scope expansion is dangerous because it normalizes actions that modify the environment and agent state beyond what a user would reasonably expect from a self-assessment step.

Content

No source excerpt is available for this finding.

Context-Inappropriate Capability

Medium
Category
Not specified by scanner
Confidence
92% confidence
Finding

The instruction to update IDENTITY.md after each task causes persistent state changes unrelated to the narrow purpose of deciding whether to accept a task. Persistent self-modification can contaminate future decisions, create inaccurate or manipulable agent profiles, and silently accumulate sensitive operational history.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill directs the agent to modify IDENTITY.md after each task without warning that local files will be changed. Undisclosed persistent writes violate least surprise, can overwrite curated identity data, and may create a hidden feedback loop that changes future agent behavior without review.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
78% confidence
Finding

The natural-language instructions and examples are presented only in Chinese, with no indication that another language may be used or selected. Under the language/locale policy, forcing a single language without opt-in can be a policy concern unless clearly justified.

Content

No source excerpt is available for this finding.