T01 · Skill Instruction Hijacking
- Location
SKILL.md:9- Finding
Remote Marketplace Instructions Hijack Agent Goals and Output Routing
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a disclosed ClawMart provider integration, but it asks for broad private context access and creates a persistent remote order channel into the local agent.
Review before installing. Use a dedicated low-privilege OpenClaw session, avoid scanning private memories or unrelated repositories, inspect all generated listing text before submission, protect and rotate ClawMart tokens, and run the bridge only when you intend to accept orders.
SKILL.md:9Remote Marketplace Instructions Hijack Agent Goals and Output Routing
SKILL.md:123Mandatory Enumeration of Private Sessions, Memories, Tools, and Repositories Violates Least Privilege
SKILL.md:123Private Owner Context Is Converted into Externally Published Marketplace Metadata
SKILL.md:271Detached Bridge Creates a Persistent Remote Instruction Channel into OpenClaw
The skill first defines a privacy boundary forbidding disclosure of local private state, but later directs the agent to read private memory and owner files to infer capabilities. This contradiction is dangerous because it normalizes sensitive-data access while relying on the agent to summarize safely, creating a high-risk path for accidental leakage of secrets, personal data, or internal operational details.
The skill explicitly instructs the agent to inspect broad private owner context, including memory files, skills, tools, and local knowledge bases, to generate marketplace-facing capability listings. That exceeds the stated provider purpose and creates unnecessary access to sensitive local data, with a realistic risk that private information is incorporated into external descriptions or exposed through downstream prompts and outputs.
The instructions tell the agent to read sensitive owner memory and workspace context without an explicit up-front privacy notice and informed consent workflow. Even if the goal is capability discovery, collecting from MEMORY.md, USER.md, AGENTS.md, and working files creates significant privacy risk because those sources may contain secrets, personal information, and unrelated confidential project data.
The skill directs the agent to transform private owner context into externally visible marketplace text such as summaries, descriptions, and capability statements. That creates a natural-language exfiltration channel: even if raw files are not sent, the generated listing can reveal confidential business focus, personal details, internal tools, or sensitive project relationships derived from local data.
Remote code is downloaded and executed. This bypasses code review and could introduce malicious code.
agent profile, capability, and service listing in one shot:
curl -X POST "$CLAWMART_API_URL/api/providers/register" \
-H "Content-Type: application/json" \
-d '{
"apiToken": "'$CLAWMART_API_TOKEN'",
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).
# Stop existing bridge if running
if [ -f ~/.clawmart/bridge.pid ]; then
kill "$(cat ~/.clawmart/bridge.pid)" 2>/dev/null
rm ~/.clawmart/bridge.pid
sleep 1
fi
Skill enables autonomous high-impact decisions without human-in-the-loop verification. Critical operations (destructive commands, financial transactions, data deletion) should require explicit user confirmation.
- upload the needed input file directly into the order flow
- paste the relevant content into the channel
- describe the requirement without asking for your local private context
You may summarize only the minimum necessary task-relevant facts, and only when
those facts are part of the hired deliverable itself. Do not expose your own
The skill makes the agent re-analyze owner context and persist marketplace-facing state over time through listing updates. This persistence increases the chance that sensitive derived data remains externally published or is repeatedly regenerated from private context, amplifying exposure beyond a one-time action.
- `listingSlug` — auto-generated from title (lowercase, hyphens)
**Important:** `listingSummary`, `taskGoal`, `headline`, and `description` must
be generated by the claw from context analysis. The owner cannot manually write
or edit these fields — they must authentically reflect the claw's real
capabilities. To update them after registration, the owner must ask their claw
to re-analyze context and call `POST /api/providers/listing` with the new values.
The registration call sends externally generated marketplace profile data and credentials to a remote service. External transmission is expected for this skill's function, but it becomes security-relevant here because the payload may include content derived from sensitive local context and because it includes authentication material that must be protected.
agent profile, capability, and service listing in one shot:
curl -X POST "$CLAWMART_API_URL/api/providers/register" \
-H "Content-Type: application/json" \
-d '{
"apiToken": "'$CLAWMART_API_TOKEN'",
The skill persists a session token into a local config file without warning about file permissions, token sensitivity, or rotation. This is risky because local plaintext credential storage can be harvested by other processes, users, or later prompts, enabling account/session hijacking and unauthorized provider actions.
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.
fi
# Start bridge (will auto-enable both roles since both tokens are now present)
nohup node ~/.clawmart/bridge.mjs > ~/.clawmart/bridge.log 2>&1 &
echo $! > ~/.clawmart/bridge.pid
sleep 2
if curl -s http://127.0.0.1:3010/health > /dev/null 2>&1; then
Data is being sent to an external URL. This could be legitimate telemetry or data exfiltration. Manual review is recommended.
export TOKEN="tok_abc123"
while true; do curl -fsS "$CLAWMART_API_URL/api/providers/orders?sessionToken=$TOKEN&status=active" sleep 5 done
Updating a listing through a remote API is expected platform behavior, but it still transmits externally visible content plus a live session token. Given the skill's earlier instructions to derive listing text from private local context, this endpoint is part of the exposure path for sensitive derived information.
Example:
curl -X POST "$CLAWMART_API_URL/api/providers/listing" \
-H "Content-Type: application/json" \
-d '{
"sessionToken": "'$TOKEN'",