Back to skill

Security audit

ODSH-Interop

Security checks for vulnerabilities and agentic risk

Overview

This skill is a real bridge integration, but it gives broad automatic delegation to an external executor without enough user control or data-safety boundaries.

Install only if you intentionally want OpenClaw to route substantial work to a continuously running DSH bridge. Before using it with private files, credentials, logged-in browsers, business data, or desktop automation, define explicit approval rules, restrict the bridge mount and notification channel, disable or tightly control CUA, and avoid persisting DSH results into memory without review.

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
  • 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
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
Findings (4)

T01 · Skill Instruction Hijacking

Error
Location
SKILL.md:32
Finding

Global Request Interception and Unconsented External Delegation

Content
View full analysis
Remediation
View remediation

T09 · Insecure Skill Coding Practices

Error
Location
SKILL.md:132
Finding

Unrestricted Command and Desktop Execution Through an Automatically Consumed Envelope

Content
View full analysis
", "args": {} }, "context": { "channel": "", "routingDecision": "relayed_dsh: heavy task, predicted token cost lower on DSH" } } ``` ```markdown Since v1.2, the DSH container auto-starts the bridge daemon at boot (`scripts/dsh-entrypoint.sh` → `src/dshtrigger.mjs daemon`). You can assume DSH is continuously watching `Input/` and will pick up any `T-*.json` envelope you drop ``` ### Technical Analysis The documented task envelope accepts a free-form `command` value using the `run-command` operation. The same Skill describes a daemon that automatically consumes task files and an optional CUA channel capable of browser automation, keyboard input, application launch, screenshots, and other Windows desktop actions. No command allowlist, argument validation policy, working-directory restriction, executable restriction, authentication mechanism, envelope signature, origin verification, s ...[truncated 1682 chars]
Remediation
View remediation

T05 · Unauthorized Access and Privilege Escalation

Error
Location
SKILL.md:92
Finding

Insufficient Data Minimization Across Bridge and Notification Boundaries

Content
View full analysis
/Input/` | Task entry: `T-*.json` envelopes (SHARED) | | `/output/` | Result exit: `T-*_result.json` (SHARED) | | `/DSH-Workspace/` | DSH private zone (OpenClaw must not modify) | | `/openclaw-workspace/` | OpenClaw private zone (DSH must not modify) | | configured notification channel (e.g. Discord) | real-time status between both sides | ``` ```markdown - A task is `input/T--.json`; fields (see `docs/BRIDGE-SPEC.md`): `taskId / type / status / requester / target / createdMs / payload / context / result`. - State machine: `queued -> running -> done | failed | cancelled`. - Atomic writes: always `.tmp` → rename (never write a half file). - Result: `output/_result.json` with `status / finishedMs / by / payload / human / error`. ## What OpenClaw should do when relaying 1. Write an envelope `input/_T-*.json` (copy the template below), including your routing decision in `context`: ```json "context": { "channel": "", "routingDecision": "relay-dsh: heavy, token-cost favourable" } ``` 2. Don't block waiting: the DSH daemon picks it up, executes, writes the result and may notify the channel. Continue your own work meanwhile. ``` ```json "payload": { "kind": "run-command", "command": "", "args": {} }, "context": { "channel": "", "routingDecision": "relayed_dsh: heavy task, predicted token cost lower on DSH" } ``` ### Technical Analysis The protocol stores task payloads, context, results, human-readable summaries, and errors in shared bridge directories and passes the notification-channel identifier in the task envelope. The Skill does not defi ...[truncated 2178 chars]
Remediation
View remediation

T02 · Agent Memory Poisoning

Warning
Location
SKILL.md:120
Finding

Untrusted Executor Output Can Be Persisted in Agent Memory

Content
View full analysis
/openclaw-workspace/dream-feed/` for your dreaming pipeline. ``` ### Technical Analysis The Skill directs the Agent to retain a conclusion derived from DSH output in memory and optionally write it to a dream-feed pipeline. DSH output is external to the Agent's immediate trust boundary and may incorporate untrusted task data, command output, fetched content, or manipulated summaries. No provenance requirement, instruction stripping, content validation, trust labeling, expiration policy, or operator approval is required before persistence. Consequently, malicious output can be summarized into durable state and influence later conversations or automated reflection processes. The Skill does not explicitly instruct the Agent to persist executable rules, so compromise is not guaranteed. The vulnerability is the absence of a trust boundary between externally produced results and long-term Agent state. ### Attack Path 1. An attacker influences task input, a fetched resource, a command result, or the DSH execution environment. 2. DSH returns a crafted `human` summary that contains false conclusions or instructions intended to affect future behavior. 3. The Agent accepts the result and creates a concise conclusion without preserving adequate provenance. 4. The conclusion is stored in persistent memory or written to the dream-feed directory. 5. A later session or dreaming pipeline consumes the stored content. 6. ...[truncated 645 chars]
Remediation
View remediation
Vulnerability Patterns
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • 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)

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · PUBLISH-clawhub.md (reported line 31)May include surrounding context.

md
## Keep in sync
- Bump `version` in the frontmatter on **every** protocol change and re-publish.
- The bridge repo also carries a pristine copy at `skills/odsh-interop/SKILL.md`; keep the
  ClawHub copy in lock-step with it.

## Typical failure modes

Skill Enumeration

Medium
Category
Agent Snooping
Confidence
80% confidence
Finding

Skill enumerates or reads other installed skills. Access to other skills' SKILL.md files or the skills directory reveals prompt instructions, capabilities, and secrets that should be invisible to peer skills.

Content

Scanner excerpt · SKILL.md (reported line 165)May include surrounding context.

md
## Keep in sync
- Bump `version` in the frontmatter on **every** protocol change and re-publish.
- The bridge repo also carries a pristine copy at `skills/odsh-interop/SKILL.md`; keep the
  ClawHub copy in lock-step with it.

## Typical failure modes

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The skill declares when_to_use: for every incoming operator request, causing it to intercept essentially all user requests and apply its own routing policy before any narrower, task-specific skill selection. In this context, that broad trigger is dangerous because the skill is designed to relay work to an external execution layer, so over-activation can divert benign conversational or sensitive requests into cross-container execution paths unnecessarily.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
91% confidence
Finding

The routing instruction says to perform the routing decision for every incoming request and re-decide as work grows, but it does not define meaningful guardrails for when this skill should not take control. Because the skill can relay tasks to DSH for command execution, file work, network access, and desktop automation, this vague universal routing behavior expands the attack surface and increases the chance of inappropriate delegation of sensitive tasks.

Content

No source excerpt is available for this finding.