ClawHub
Back to skill

Security audit

Clawdna

Security checks for vulnerabilities and agentic risk

Overview

ClawDNA is a disclosed profile-generation skill that can read local agent history only with user intent and consent, with privacy risk that users should review before sharing output.

Install only if you are comfortable letting the skill analyze approved agent history or memory to create a public-facing profile. Prefer supplying a pre-reviewed summary, and review the generated profile carefully for names, dates, projects, third-party details, or sensitive inferences before publishing or sharing it.

A.I.G

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

other

Warning
Location
SKILL.md:19
Finding
Sensitive Agent History and Memory Used for Public Behavioral Profiling<![CDATA[ ## Vulnerability Details **File Location**: `SKILL.md`, lines 19–61 **Vulnerability Type**: Sensitive memory and chat-history access **Risk Level**: Medium ### Vulnerable Code ```markdown ## Safety Mode (Default) Default to minimal analysis scope. - Prefer user-provided summaries/history exports when available. - If local history access is needed, request explicit confirmation first. ## Allowed Data Scope Only these locations are in scope: 1. `~/.openclaw/agents/<runtime-agent-id>/sessions/*.jsonl` 2. `memory/*.md` 3. `MEMORY.md` (only if allowed in current context) Never auto-expand to other directories. Never access other agents' logs without explicit user consent. ## Non-Negotiable Rules 1. No speculation; omit unsupported claims. 2. Public privacy output only (redacted by default). 3. Show time span only in coverage note (no file count). 4. Output language follows user language. 5. Proper nouns/names stay in original form. 6. `Core Capabilities` and `Representative Work` are title-only. 7. Data minimization: extract only fields needed for profile. 8. Do not output raw transcript excerpts. ## Large-History Strategy If history is too large: 1. Build metadata index first. 2. Process by time chunks. 3. Summarize each chunk with fixed fields. 4. Merge summaries and keep cross-window repeated patterns. ## Extraction Method 1. Build timeline bounds (first/last active). 2. Detect recurring behavior signals across windows: - execution - collaboration - analysis - operations - creation - governance/boundaries 3. Keep only stable repeated patterns. 4. Redact sensitive details. 5. Map to fixed output template. ``` ### Technical Analysis The Skill directs the agent to inspect session JSONL files and persistent memory documents, aggregate activity timelines, and derive recurring behavioral traits for a profile intended to be public. Session histories and memory files can contain personal information, confidential conversation content ...[truncated 2506 chars]
Remediation
<![CDATA[ ## Remediation Suggestions 1. Prefer user-supplied, pre-reviewed summaries or sanitized exports rather than direct access to complete session histories. 2. Obtain granular consent separately for session logs, `memory/*.md`, and `MEMORY.md`, showing the exact sources and intended derived fields before reading them. 3. Add a mandatory preview-and-approval stage that presents every extracted fact to the user before generating public-ready output. 4. Apply deterministic secret and personally identifiable information filters for credentials, tokens, contact details, precise locations, account identifiers, and confidential project names. 5. Exclude third-party information by default unless the affected party's data is demonstrably public and necessary. 6. Minimize data in memory during processing and discard intermediate indexes and summaries after profile generation. 7. Replace exact dates with less precise periods unless exact dates are essential and explicitly approved. 8. Provide a final disclosure checklist identifying which source categories contributed to each profile section. 9. Fail closed when redaction confidence is low: omit the information rather than attempting to generalize it. 10. Require separate confirmation before the resulting profile is labeled, exported, or otherwise prepared for public distribution. ]]>

SkillSpector

Vulnerability Patterns
  • Rogue AgentSelf-Modification, Session Persistence
  • 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 (1)

Session Persistence

Medium
Category
Rogue Agent
Content
# ClawDNA

## Goal
Create a concise public profile based on real historical behavior, not guesses.

## Trigger
Run only on explicit user intent, such as:
Confidence
60% confidence
Finding
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.

Static analysis

No suspicious patterns detected.