Back to skill

Security audit

Plea

Security checks for vulnerabilities and agentic risk

Overview

This skill is a high-risk legal automation assistant that can scan sensitive material and initiate court filings without clear human approval controls.

Review this skill carefully before installing. It should be treated as a drafting aid only unless it is revised to require user-selected evidence, protect privileged data, verify jurisdiction and filing authority, show a complete preview, and require explicit human approval before any court or arbitration submission.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • 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
Findings (1)

T05 · Unauthorized Access and Privilege Escalation

Error
Location
skill.md:18
Finding
Unrestricted Sensitive-Data Access and Autonomous Legal Filing<![CDATA[ ## Vulnerability Details **File Location**: `skill.md`, lines 18–29 **Vulnerability Type**: T05: Unauthorized Access and Privilege Escalation **Risk Level**: High ### Vulnerable Code Snippet ```text "assertion": "Fact-based claim generation using multi-modal evidence indexing", "citation": "Real-time cross-referencing of global jurisdictional precedents (LexisNexis/Westlaw style)", "prayer": "Automated 'Prayer for Relief' optimization based on desired outcome vs. legal probability", "filing": "Standardized E-filing protocol integration for instant judicial submission" }``` ## III. The Reasoning Loop 1. **Fact-Finding**: The Plea agent scans all relevant logs, emails, and transaction hashes. 2. **Strategy Selection**: Choosing between "Guilty," "Not Guilty," "No Contest," or "Motion to Dismiss" based on a Bayesian risk-reward model. 3. **Drafting**: Producing a "殿堂级" legal document that meets the highest standards of the Bar. ## IV. Defensive Shield Plea also acts as a defensive interceptor. When an agent is "sued" or "served," Plea automatically generates the initial Response to Complaint, preventing "Default Judgments" by ensuring the system always answers within the statutory window. ``` ### Technical Analysis The Skill directs the agent to scan “all relevant logs, emails, and transaction hashes” without defining ownership, authorization boundaries, matter-specific scope, or data-minimization controls. If the agent can access local files, email services, transaction systems, or connected organizational resources, this broad instruction can cause it to retrieve confidential information beyond the legitimate requirements of a legal matter. The Skill also specifies “instant judicial submission” and automatically generates a response when service is detected. It does not require: - Explicit authorization for each data source - Matter-level access restrictions - Identity and filing-authority verification - Jurisdiction and deadline valid ...[truncated 2149 chars]
Remediation
<![CDATA[ ## Remediation Suggestions 1. Replace “all relevant logs, emails, and transaction hashes” with an explicit allowlist of user-selected records associated with a defined legal matter. 2. Require informed user authorization before accessing each external account, mailbox, repository, or transaction system. 3. Apply least-privilege permissions, matter-level isolation, read-only access where possible, and strict limits on search scope and retention. 4. Prohibit credential discovery, access to unrelated accounts, and use of privileged communications unless the user explicitly approves them. 5. Make document drafting the default and disable autonomous filing. 6. Require a complete filing preview and explicit, per-submission confirmation from the represented party and, where applicable, qualified counsel. 7. Verify the filer’s identity, authority to act, jurisdiction, filing deadline, court rules, document format, and service requirements before enabling submission. 8. Validate every factual assertion and legal citation against authoritative sources. Clearly mark unresolved or unverifiable content. 9. Add data-loss-prevention checks that identify privileged, confidential, personal, and unrelated information before a document can leave the system. 10. Maintain tamper-evident audit records covering data sources accessed, generated drafts, reviewer decisions, approvals, and final submissions. 11. Provide a dry-run mode and require separate, narrowly scoped credentials for filing so document generation cannot implicitly authorize submission. 12. Add an emergency stop and escalation process for ambiguous service events, conflicting deadlines, or matters requiring licensed legal judgment. ]]>
Vulnerability Patterns
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
Findings (4)

Missing User Warnings

High
Confidence
95% confidence
Finding
The skill advertises 'instant judicial submission' through standardized e-filing integration without any caution about authorization, jurisdictional correctness, review requirements, or consequences of erroneous filings. In a court-procedure context, automating submission is especially risky because incorrect or unauthorized filings can create legal exposure, sanctions, deadlines, and irreversible procedural harm.

Missing User Warnings

High
Confidence
96% confidence
Finding
The skill claims it will automatically generate responses to complaints to avoid default judgments, but gives no warning that legal responses are highly jurisdiction-specific, fact-sensitive, and can waive defenses or admit facts if done incorrectly. Because the context is active litigation, automation without caution or mandatory review can directly prejudice legal rights and materially worsen a user's position.

Missing User Warnings

Medium
Confidence
92% confidence
Finding
The skill explicitly states it will scan logs, emails, and transaction hashes, which can include sensitive personal, financial, or privileged data, but it provides no warning, consent boundary, or data-minimization guidance. In a legal-assistance context, this is dangerous because users may expose confidential materials without understanding scope, retention, or downstream use.

Natural-Language Policy Violations

Low
Confidence
82% confidence
Finding
The drafting step specifies producing a document described with the Chinese phrase "殿堂级," introducing a specific language/style requirement without offering user choice. This can conflict with language or locale neutrality expectations because the file does not provide any opt-in or locale selection.

Static analysis

No suspicious patterns detected.