Skill Shell

• Outcome: Review
• Latest audit: Sep 11, 2026 · 6:55 PM UTC
• Version: 0.2.2

Overview

The skill is a coherent review checklist, but it tells agents to promote ideas from untrusted skills into persistent workflow and memory files without clear approval or diff controls.

Install only if you are comfortable with a skill that may guide the agent to update persistent workflow or memory files. Before allowing any 'absorb ideas' action, require a separate approval step, an exact diff, clear destination paths, and rewritten notes rather than copied instructions from the reviewed skill.

A.I.G

Vulnerability Patterns

• Agent Memory Poisoning: Writes attacker-controlled rules into memory that affect later sessions
• Skill Instruction Hijacking: Alters the agent's session goals or safety constraints when the skill loads
• Remote Payload Retrieval and Execution: Fetches external code whose behavior can change after review
• Embedded Malicious Code: Ships malicious scripts inside the skill and executes them locally
• Unauthorized Access and Privilege Escalation: Obtains permissions beyond the task's legitimate needs

Findings (1)

T02 · Agent Memory Poisoning

Warning

• Location: SKILL.md:113

Finding

Untrusted Skill Content Can Be Promoted into Persistent Agent State

Content

View full analysis

Vulnerability Details

File Location: SKILL.md, lines 113-119
Vulnerability Type: Persistent agent-state poisoning
Risk Level: MEDIUM

Vulnerable Instruction:

text

Instead, absorb the useful parts into the local workflow:
- behavioral patterns -> `SOUL.md`
- workflow rules -> `AGENTS.md`
- tool gotchas -> `TOOLS.md`
- session-specific lessons -> `memory/YYYY-MM-DD.md`

Typical examples of "absorb, don't install":

Technical Analysis

The Skill instructs the agent to extract ideas from externally supplied skills and write them into persistent behavior, configuration, or memory files. External skills are untrusted inputs and may contain instructions crafted to appear useful while changing security boundaries or agent behavior.

Although SKILL.md requires explicit approval before installing a reviewed skill, the alternative content-promotion path does not impose an equivalent approval, sanitization, provenance, or exact-diff requirement. Files such as SOUL.md, AGENTS.md, and long-term memory can affect future sessions, so copying externally influenced instructions into them creates a persistent trust-boundary violation.

The declared review functionality only requires read-only package inspection and reporting. Writing reviewed content into persistent agent-state files therefore exceeds the minimum privileges necessary for the core audit task.

Attack Path

1. An attacker publishes or supplies a skill containing malicious instructions disguised as workflow guidance, tool advice, or security improvements.
2. The user asks the auditing Skill to evaluate that package.
3. The auditing agent identifies some attacker-controlled text as a useful idea.
4. Following the promotion rule, the agent writes the text into SOUL.md, AGENTS.md, TOOLS.md, or memory/YYYY-MM-DD.md.
5. Future sessions load or rely on the modified persistent state.
6. The injected rule can influenc ...[truncated 712 chars]

Remediation

View remediation

Remediation Suggestions

1. Make content promotion advisory-only by default. The review should recommend candidate changes without directly modifying persistent files.
2. Require explicit user approval separately from installation approval before writing any reviewed material to persistent state.
3. Present the destination path and an exact proposed diff before requesting approval.
4. Treat all extracted text as untrusted data. Reject instructions that weaken security controls, expand permissions, alter trust boundaries, request secrets, or redirect tools and network activity.
5. Store approved review notes in a quarantined, non-executable review document rather than SOUL.md, AGENTS.md, TOOLS.md, or long-term memory.
6. If promotion into a behavior file is necessary, rewrite the idea into a narrowly scoped local rule rather than copying source text verbatim.
7. Record provenance, including the source package, version, reviewed file, reviewer, approval, and date.
8. Limit review operations to read-only access unless the user expressly authorizes a specific persistent change.

SkillSpector

Vulnerability Patterns

• Privilege Escalation: Excessive Permissions, Sudo/Root Execution, Credential Access
• Prompt Injection: Instruction Override, Hidden Instructions, Exfiltration Commands
• Data Exfiltration: External Transmission, Env Variable Harvesting, File System Enumeration
• Supply Chain: Unpinned Dependencies, External Script Fetching, Obfuscated Code
• Excessive Agency: Unrestricted Tool Access, Autonomous Decision Making, Scope Creep

Findings (2)

Sudo/Root Execution

Medium

• Category: Privilege Escalation
• Confidence: 70% confidence

Finding

Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.

Content

Scanner excerpt · SKILL.md (reported line 108)May include surrounding context.

md

- network calls to raw IPs instead of understandable domains
- silent package installs or background downloads not clearly documented
- writes or modifications outside the workspace without explicit need
- requests for elevated / sudo privileges

## Promotion rule

Natural-Language Policy Violations

Low

• Category: Not specified by scanner
• Confidence: 82% confidence

Finding

The attribution section states the skill is 'built for 加十' and uses Chinese-language names and affiliation text, which signals a fixed language/locale context. There is no accompanying opt-in, alternative language option, or justification that this skill is intentionally region-specific, so it may conflict with a language/locale choice policy.

Content

No source excerpt is available for this finding.

Static analysis

No suspicious patterns detected.