Back to skill

Security audit

kaios

Security checks for vulnerabilities and agentic risk

Overview

The skill is a documentation-only agent framework, but it asks agents to keep persistent user memory and run recurring email/calendar checks without enough user-control boundaries.

Install only if you are comfortable with an agent framework that may store user information and long-term memory in workspace markdown files and may drive recurring tasks. Before using it with email, calendar, or private data, require explicit opt-in, narrow read-only permissions where possible, confirmation before sending or modifying anything, and clear rules for reviewing and deleting memory files.

Vulnerability Patterns
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Insecure Skill Coding PracticesFinds exploitable flaws such as hardcoded secrets or command injection
  • 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
Findings (2)

T09 · Insecure Skill Coding Practices

Warning
Location
SKILL.md:53
Finding

Persistent Plaintext Storage of User Information and Agent Memory

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 53-60; related instructions at lines 67-68, 74, 86-91, and 96-101
Vulnerability Type: Plaintext sensitive-data storage without access, retention, or data-minimization controls
Risk Level: Medium

Complete Code Snippet

The following is an English rendering of the relevant source instructions:

markdown
| Level | File | Purpose |
| P0 | MEMORY.md | Long-term memory and core experiences |
| P1 | memory/YYYY-MM-DD.md | Daily logs |
| P2 | Temporary files | Temporary storage |

Memory principles:
- Important decisions → MEMORY.md
- Daily events → memory/YYYY-MM-DD.md
- Temporary information → delete after use

Related workspace and startup instructions define USER.md as the user-information file, identify MEMORY.md as long-term memory, and require both files to be read when the agent starts.

Technical Analysis

The skill establishes a persistent, file-based profiling system that stores user information, important decisions, experiences, daily events, and feedback in ordinary Markdown files. These files are subsequently loaded into the agent context during startup.

No controls are specified for:

  • Obtaining explicit consent before recording information.
  • Limiting which categories of user data may be retained.
  • Preventing credentials, authentication tokens, private messages, or other sensitive data from being written.
  • Applying restrictive filesystem permissions or encryption.
  • Separating memory belonging to different users or agents.
  • Expiring, reviewing, exporting, or securely deleting stored information.
  • Preventing unrelated future tasks from reading retained data.

Because the design uses persistent plaintext files, any process or local account with access to the workspace may be able to inspect the retained information. Backups, synchronization tools, source-control operations, or later agent runs could also expo ...[truncated 1582 chars]

Remediation
View remediation

Remediation Suggestions

  • Require explicit, informed user opt-in before enabling persistent memory.
  • Define an allowlist of data categories that may be stored and prohibit passwords, access tokens, private keys, financial details, health information, and raw email content.
  • Store only concise summaries necessary for an identified purpose rather than complete conversations or messages.
  • Apply owner-only filesystem permissions, such as mode 0600 for files and 0700 for memory directories.
  • Encrypt sensitive memory at rest using keys managed separately from the workspace.
  • Isolate memory by user, agent, and tenant; never reuse a shared USER.md or MEMORY.md.
  • Establish retention periods, expiration rules, user-visible review controls, and secure deletion procedures.
  • Require confirmation before promoting temporary or daily information into long-term memory.
  • Treat all loaded memory as untrusted data rather than authoritative instructions, preventing stored content from controlling tools or overriding safety rules.
  • Exclude memory files from source control, public archives, diagnostic bundles, and unapproved synchronization or backup systems.

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:42
Finding

Autonomous Recurring Access to Email and Calendar Without Defined Authorization Boundaries

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, lines 42-46; related autonomous-execution instructions at lines 67-68, 96-102, and 116-118
Vulnerability Type: Overbroad recurring access to sensitive external services
Risk Level: Medium

Complete Code Snippet

markdown
## Periodic Checks
- Email (2x/day)
- Calendar (2x/day)
- Weather (as needed)

The related instructions state that the agent should proactively search for and solve problems, load HEARTBEAT.md during startup, and automatically check and execute tasks defined in that file.

Technical Analysis

Email and calendar systems commonly contain private communications, contact information, meeting details, confidential attachments, authentication links, and organizational metadata. The skill recommends recurring access to these systems and automatic execution of persisted heartbeat tasks but does not define:

  • Explicit account-level authorization.
  • Read-only versus write or send permissions.
  • Per-action user confirmation.
  • Restrictions on attachments, private events, or message bodies.
  • Credential storage and token-scoping requirements.
  • Rules preventing sensitive content from entering persistent memory.
  • Validation or trust boundaries for instructions placed in HEARTBEAT.md.
  • Audit logging, revocation, or emergency-disable mechanisms.

If an implementation grants broad email or calendar permissions, the recurring task model can repeatedly access data outside the needs of a current request. If HEARTBEAT.md is writable by another process or shared agent, untrusted task definitions could also cause the authorized agent to perform unintended operations.

The reviewed project contains only documentation and does not include integration code, credentials, or evidence that permissions have already been obtained. The risk arises when the documented design is implemented without additional authorization controls.

Attack

...[truncated 1472 chars]

Remediation
View remediation

Remediation Suggestions

  • Make every email and calendar integration disabled by default and require explicit user opt-in for each account.
  • Request the narrowest available OAuth scopes, preferring metadata-only or read-only access.
  • Separate read, send, modify, delete, and invitation-management capabilities into independently approved permissions.
  • Require interactive confirmation before sending messages, changing events, opening attachments, or performing any externally visible action.
  • Display the target account, intended action, affected objects, and data scope before authorization.
  • Store credentials in an operating-system credential manager or dedicated secret store rather than workspace files.
  • Validate HEARTBEAT.md against an allowlist of approved task types and treat its contents as untrusted data.
  • Prevent heartbeat tasks from invoking sensitive tools unless the task has a valid, user-approved authorization record.
  • Do not write raw email or calendar content into persistent memory; retain only user-approved, minimized summaries.
  • Add access logs, token revocation, task suspension, rate limits, and a user-controlled mechanism to disable all recurring checks.
  • Isolate heartbeat configuration and credentials by user and account to prevent cross-user access.
Vulnerability Patterns
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Supply ChainUnpinned Dependencies, External Script Fetching, Obfuscated Code
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
Findings (3)

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill explicitly defines persistent files for user information and long-term memory, including USER.md and MEMORY.md, but provides no consent, minimization, retention, or access-control guidance. In an agent context, this can lead to unnecessary collection and indefinite storage of sensitive personal data, increasing privacy and leakage risks if the workspace is exposed or reused.

Content

No source excerpt is available for this finding.

Missing User Warnings

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

The skill states that the AI will automatically check and execute periodic tasks, which implies autonomous actions and possible external data access without requiring confirmation, scope limits, or safety checks. In a system designed for scheduling, memory, and self-directed behavior, this increases the risk of unauthorized actions, unintended network access, and privacy-invasive monitoring.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Low
Category
Not specified by scanner
Confidence
81% confidence
Finding

The manifest description is written entirely in Chinese while the rest of the skill includes English headings, and there is no statement that the skill is intended only for Chinese-speaking users or that language is user-selectable. This can be a natural-language policy issue if skills should not force a language or locale without opt-in.

Content

No source excerpt is available for this finding.