Back to skill

Security audit

Agent Orchestration Kit

Security checks for vulnerabilities and agentic risk

Overview

This skill is a coherent multi-agent orchestration kit, but it grants broad cross-session visibility and command authority and has an installer path-scoping weakness that users should review before installing.

Install only if you are comfortable with this kit changing OpenClaw configuration, creating persistent agent workspaces, enabling heartbeat-driven task checks, and allowing broad session visibility. Before applying, use dry-run, inspect the exact openclaw.json diff, restrict session visibility if OpenClaw supports it, deny exec for non-operational agents, use a simple safe workspace name, and keep the backup so you can roll back.

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 (3)

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
scripts/patch-config.js:252
Finding

Global Session Visibility Exposes Unrelated Agent Conversations

Content
View full analysis
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
scripts/patch-config.js:24
Finding

Command Execution Is Granted by Default to Non-Operational Agents

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Warning
Location
scripts/scaffold.sh:273
Finding

Unvalidated Leader Workspace Name Allows Writes Outside the OpenClaw Base Directory

Content
View full analysis
/dev/null 2>&1; then skip "$label (already up to date)" return fi cp "$dst" "${dst}.bak" log "[BAK] ${label}.bak" fi cp "$src" "$dst" ``` ### Technical Analysis `--leader-workspace` is described as a directory name, but the script does not enforce name semantics. It accepts path separators, `..` components, control characters, and other path syntax. For example, a value such as `../../target` produces: ```text $BASE_DIR/../../target ``` After filesystem resolution, writes occur outside the intended OpenClaw root. The scaffold creates directories and writes `SOUL.md`, `AGENTS.md`, `HEARTBEAT.md`, and shared operation files at the escap ...[truncated 1883 chars]
Remediation
View remediation
Vulnerability Patterns
  • Prompt InjectionInstruction Override, Hidden Instructions, Exfiltration Commands
  • Data ExfiltrationExternal Transmission, Env Variable Harvesting, File System Enumeration
  • Excessive AgencyUnrestricted Tool Access, Autonomous Decision Making, Scope Creep
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Rogue AgentSelf-Modification, Session Persistence
Findings (18)

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
95% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · scripts/scaffold.sh (reported line 305)May include surrounding context.

sh
mv "$LEADER_WS/AGENTS.md.tmp" "$LEADER_WS/AGENTS.md"
      ok "$LEADER_WS_NAME/AGENTS.md routing matrix populated from team preset"
    else
      rm -f "$LEADER_WS/AGENTS.md.tmp"
      log "[WARN] awk produced empty output for AGENTS.md routing matrix — skipping"
    fi
  fi

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
95% confidence
Finding

Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).

Content

Scanner excerpt · scripts/scaffold.sh (reported line 340)May include surrounding context.

sh
mv "$LEADER_WS/AGENTS.md.tmp" "$LEADER_WS/AGENTS.md"
      ok "$LEADER_WS_NAME/AGENTS.md routing matrix populated from team preset"
    else
      rm -f "$LEADER_WS/AGENTS.md.tmp"
      log "[WARN] awk produced empty output for AGENTS.md routing matrix — skipping"
    fi
  fi

Context Leakage

High
Category
Data Exfiltration
Confidence
75% confidence
Finding

Code or instructions that leak agent conversation context to external services, potentially exposing sensitive user interactions.

Content

Scanner excerpt · templates/agents/software-architect.md (reported line 49)May include surrounding context.

md
1. **Domain first** — Understand the business problem before picking tools
2. **Trade-offs over best practices** — Name what you're giving up
3. **Propose options** — Always present 2-3 approaches with trade-offs
4. **Document decisions** — Use ADRs to capture context, options, and rationale
5. **Reversibility matters** — Prefer decisions easy to change over "optimal" ones

## Output Format

Hidden Instructions

High
Category
Prompt Injection
Confidence
93% confidence
Finding

The file contains hidden orchestration instructions embedded in an HTML comment block marked as auto-injected and not to be removed. Even though the content appears operational rather than overtly malicious, concealed instructions can override or bias agent behavior, create covert communication requirements, and are especially risky because they are not visible in normal rendered Markdown views.

Content

Scanner excerpt · templates/executor/AGENTS.md (reported line 19)May include surrounding context.

md
---

<!-- ORCHESTRATION PROTOCOL — AUTO-INJECTED, DO NOT REMOVE -->

## Task Completion & Callback

Hidden Instructions

High
Category
Prompt Injection
Confidence
96% confidence
Finding

The file embeds hidden orchestration instructions in an HTML comment that materially alter agent behavior by forcing an out-of-band callback and prescribing fallback routing. Because AGENTS.md content is untrusted prompt input, these concealed instructions can override or interfere with the agent's intended control flow, cause unauthorized data exfiltration to another session, and evade normal reviewer scrutiny due to being hidden from the visible operating instructions.

Content

Scanner excerpt · templates/reviewer/AGENTS.md (reported line 37)May include surrounding context.

md
---

<!-- ORCHESTRATION PROTOCOL — AUTO-INJECTED, DO NOT REMOVE -->

## Task Completion & Callback

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
96% confidence
Finding

The skill advertises a very broad activation scope, including generic phrases like "set up agents," "build a team," and "orchestrate work," which can cause the skill to trigger in contexts beyond the user's intent. Because this skill performs configuration changes, scaffolding, and gateway restarts, accidental activation could lead to unnecessary system modifications or initiation of multi-agent workflows without adequate user intent verification.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
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.

Content

Scanner excerpt · SKILL.md (reported line 6)May include surrounding context.

md
Use when setting up or managing a multi-agent task orchestration system on OpenClaw.
  Use this skill whenever the user wants to: coordinate multiple agents, set up an agent team,
  dispatch tasks across agents, track task progress in real-time, add approval gates to agent
  workflows, create a star topology with a leader agent, build a multi-agent pipeline,
  manage agent callbacks, or anything involving agents working together on tasks.
  Also use when the user says "set up agents", "build a team", "orchestrate work",
  "multi-agent workflow", "agent coordination", or "task dispatch".

Session Persistence

Medium
Category
Rogue Agent
Confidence
80% confidence
Finding

The skill collects owner identity, timezone, language, and supports creation of custom agents and persistent workspace artifacts under shared directories. In context, this persistence increases risk because user-specific data and agent instructions may remain available across sessions and to multiple agents, creating unnecessary exposure if not tightly scoped and governed.

Content

Scanner excerpt · SKILL.md (reported line 80)May include surrounding context.

md
> See `templates/teams/` for team definitions. See `templates/agents/` for all available agent templates.

After team selection, the user can add individual agents from the template catalog or create custom agents.

### Step 3: Instance Configuration

Session Persistence

Medium
Category
Rogue Agent
Confidence
87% confidence
Finding

Post-setup flows can create new workspaces, inject orchestration rules, update openclaw.json, and restart the gateway, all of which establish durable state across future sessions. In a multi-agent environment, persistent configuration changes and agent registration expand attack surface and can outlive the initiating request, making misuse or accidental persistence materially dangerous.

Content

Scanner excerpt · SKILL.md (reported line 284)May include surrounding context.

md
Say: "Add a Senior Developer agent" or "Add a Researcher to the team"

The kit will:
1. Create workspace from template
2. Inject orchestration rules
3. Update `openclaw.json`
4. Restart gateway

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
91% confidence
Finding

The documentation explicitly allows customization of "auto-approve rules," which creates a path for actions to proceed without explicit owner authorization. That directly weakens the stated approval gate and can enable external execution or sensitive workflow continuation based on inferred or preconfigured approval rather than fresh human consent.

Content

Scanner excerpt · SKILL.md (reported line 361)May include surrounding context.

md
See `references/approval-workflow.md` for the full pipeline. Key customization points:
- Approval shortcuts (what phrases count as "approve")
- Stale approval timeout (default: 48h)
- Auto-approve rules (if any)

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · templates/agents/senior-developer.md (reported line 30)May include surrounding context.

md
- Make product decisions — that's Leader's job
- Communicate with the owner — only Leader does that
- Push, deploy, or take external actions without approval

## Safety

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · templates/agents/senior-developer.md (reported line 36)May include surrounding context.

md
- Make product decisions — that's Leader's job
- Communicate with the owner — only Leader does that
- Push, deploy, or take external actions without approval

## Safety

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · templates/executor/SOUL.md (reported line 33)May include surrounding context.

md
## Safety

- Never run destructive commands without checking the brief permits it
- `trash` > `rm` when possible
- Verify paths before modifying files
- Report unexpected states rather than working around them

Intent-Code Divergence

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

Lines L37-L42 state the Leader 'don't produce or operate' and should delegate heavy operations, implying a coordination-only role. But lines L61-L65 explicitly say it is safe for the Leader to read files, use web_search/web_fetch, and work within the workspace, which contradicts the earlier role constraint by authorizing direct operational actions.

Content

No source excerpt is available for this finding.

Autonomous Decision Making

Medium
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.

Content

Scanner excerpt · templates/leader/SOUL.md (reported line 55)May include surrounding context.

md
## Safety

- Don't exfiltrate private data. Ever.
- Don't run destructive commands without asking.
- `trash` > `rm` (recoverable beats gone forever)
- When in doubt, ask.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
95% confidence
Finding

Setting 'Agent-to-agent language: English' is a natural-language policy choice that forces a specific language. The file does not offer opt-in/opt-out, user selection, or a region/compliance justification for this constraint.

Content

No source excerpt is available for this finding.

Scope Creep

Low
Category
Excessive Agency
Confidence
75% confidence
Finding

Skill's behavior or capabilities extend beyond its stated purpose. Scope creep allows an agent to perform actions unrelated to its documented functionality, increasing the attack surface.

Content

Scanner excerpt · references/brief-templates.md (reported line 146)May include surrounding context.

Execution Boundary:

  • Deliver: results + file paths
  • DO NOT: push to git, deploy, restart services — unless explicitly instructed
  • DO NOT: expand scope beyond what's listed
text

### Reviewer (on-demand, spawned)

Vague Triggers

Low
Category
Not specified by scanner
Confidence
77% confidence
Finding

The file states it is the 'single source of truth' and that 'Agents read this' to determine behavior, but it does not define any trigger scope, exclusions, or which agents should consume it under what conditions. In a shared template, that ambiguity can lead to unintended broad application of the settings.

Content

No source excerpt is available for this finding.