T05 · Unauthorized Access and Privilege Escalation
- Location
config/default.json:19- Finding
Direct Cross-Agent Memory Access Weakens Isolation Boundaries
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This memory-management skill is coherent in purpose, but it needs review because it enables cross-agent memory access and includes risky file-management guidance without enough guardrails.
Review before installing. Disable or narrow the default writing-assistant memory path unless you explicitly want this skill to read that agent's memory. Do not run missing PowerShell scripts from an unverified source, avoid ExecutionPolicy Bypass, and replace the Force deletion/move examples with preview, backup, and explicit confirmation steps.
config/default.json:19Direct Cross-Agent Memory Access Weakens Isolation Boundaries
docs/QUICK_START.md:133Recursive Forced Deletion Uses an Unvalidated Relative Path
docs/QUICK_START.md:146Troubleshooting Guidance Disables PowerShell Execution-Policy Checks
The document explicitly states that all operations are automatically logged to JSON files, but it does not clearly disclose what data is recorded, where it is stored, how long it is retained, or how users can disable it. In a memory-management skill handling agent activity and shared workspace operations, silent or poorly disclosed logging can capture sensitive prompts, file paths, project data, or user actions and create privacy and confidentiality risk.
The README restricts valid confirmations to a Chinese-language allowlist such as “是”, “允许”, “确认”, and “好的”, without indicating multilingual support or a configurable confirmation policy. In a safety-sensitive workflow like deletion or archival, language-restricted confirmation handling can cause incorrect rejection of legitimate user intent or push users toward ambiguous phrasing, increasing the risk of unsafe or failed authorization handling.
The documentation explicitly encourages path-based access to another agent's memory and presents inter-agent memory sharing as a normal workflow, but it does not define consent, access boundaries, sanitization, or privacy warnings. In a multi-agent environment, this can lead to unintended disclosure of sensitive prompts, workspace data, credentials, or user content across agent boundaries.
The document title and all explanatory content are written exclusively in Chinese, with no indication that language selection is optional or that the skill is intended only for a Chinese-speaking or region-specific context. Under the language/locale policy, forcing a specific language without user opt-in can be a natural-language policy violation.
The design principles state '无定期自动任务' and that all operations are executed on demand, implying no scheduled automation. However, the file later describes '定期检查', a '定期清理机制', and '定期维护', which contradict the earlier intent statement rather than merely adding detail.
The document asserts that logs are immutable, but only describes ordinary JSON files stored in a logs directory with no append-only, signing, ACL, WORM, or remote attestation mechanism. If operators rely on this claim, an attacker or insider with filesystem access could alter or delete audit records and undermine forensic integrity.
The quick-start includes a cleanup command that recursively deletes files with Remove-Item -Force but does not clearly warn users that it is destructive or advise verification/dry-run steps first. In operational documentation, presenting deletion commands as routine maintenance increases the chance of accidental data loss, especially for inexperienced users following copy-paste instructions.
The troubleshooting section instructs users to manually move files with Move-Item ... -Force, which can overwrite or relocate data without adequate warning about operational impact. Because this is framed as a manual recovery step, users may execute it under pressure and unintentionally lose track of files or overwrite archive contents.
The natural-language content of the skill file is entirely in Chinese, which may impose a language expectation on users without any stated opt-in or locale justification. Under the policy, forcing a specific language without user choice can be a language/locale policy issue.
The natural-language content of the skill is presented in Chinese throughout the document, which effectively forces a specific language for users. The file does not offer alternative language support or state that the skill is intended only for a Chinese-language audience.
The agent description at L25 is written in Chinese while surrounding configuration fields are in English, indicating an implicit language choice without any documented user opt-in or locale justification. The policy for this review flags language or locale constraints when they are imposed without offering a choice or explaining the regional scope.
The principle '用户完全控制' suggests operations occur only under explicit user control. Later, the tool call chain shows 'simple_logger.ps1 (自动记录)', indicating an automatic side effect that happens as part of the workflow, which conflicts with the absolute wording of full user control.
The entire skill documentation is presented only in Chinese and does not indicate that other languages are available or that Chinese is required for a justified regional context. This can be a natural-language locale policy issue when a skill effectively mandates one language without opt-in or explanation.