Security audit
Security checks for vulnerabilities and agentic risk
CogDx is a disclosed external diagnostics client that sends user-provided agent outputs and reasoning data to its API for paid analysis, with no hidden execution or persistence found.
Install only if you are comfortable sending selected agent prompts, outputs, claims, context, confidence values, reasoning traces, feedback, and a wallet/account identifier to Cerebratech's external API. Do not submit secrets, regulated data, proprietary reasoning, or private user data unless you have authorization and have redacted it first.
The skill demonstrates capabilities that use environment variables and external network access, but it does not declare any tool scope or permissions boundaries. This can cause the skill to be run with broader privileges than users expect, reducing transparency and increasing the chance of unintended data access or outbound transmission.
The examples show prompts, responses, predictions, and full reasoning traces being sent to an external service for analysis, but the skill description does not clearly warn users that potentially sensitive model inputs and internal reasoning data leave the local environment. In a diagnostic skill, this context makes the issue more dangerous because users are encouraged to upload precisely the kinds of sensitive or proprietary cognitive artifacts that may contain secrets, internal policies, personal data, or business logic.
The skill pulls a wallet identifier from an environment variable and uses it as an authentication/payment credential for a third-party service. Even if intended for billing, this expands the skill's access beyond diagnostics and creates a credential-handling path that could enable unauthorized charging, identity linkage, or unintended external account use if the skill is invoked without clear operator consent.
Multiple methods transmit agent outputs, predictions, and full reasoning traces to an external API, and the callable interfaces do not provide explicit warnings, consent hooks, or redaction controls. In an agent context, those payloads can contain sensitive prompts, proprietary data, credentials, or chain-of-thought-like internal reasoning, making unannounced export to a third party a significant confidentiality risk.
The feedback submission path goes beyond diagnostics and sends wallet, agent identity, and free-form reasoning/comments to an external endpoint to earn credits. That creates an unnecessary data-exfiltration and incentive mechanism: operators may disclose sensitive internal outputs or evaluations to accumulate credits, increasing privacy and governance risk.
The code silently sources a wallet from the environment and transmits it in an authentication header, without making credential handling obvious at the request boundary. This can surprise integrators, increase accidental secret exposure risk, and make it easier for deployments to leak account-linked identifiers to a third party without deliberate review.
get_credits queries an external /credits endpoint to retrieve account balance and spending information. Account/billing inspection is separate from calibration audits, bias detection, reasoning verification, and consensus building, and the manifest does not indicate that financial account management is part of the skill's role.
No suspicious patterns detected.
