Trust has to be inspectable.
An important conclusion needs evidence. A consequential action needs authority.
Evidence before organisational truth.
Claims reference tenant-scoped evidence such as API results, configuration, repository state, identity relationships, threat intelligence, or human confirmation. Confidence and concise reasoning accompany the claim; uncertainty remains visible.
Human context changes how an observation is understood, not whether an innate weakness exists. An intentionally public HTTPS service still needs appropriate origin protection, authentication, TLS, and operating controls.
Investigation is not remediation authority.
Approved read-only investigation may query, correlate, form hypotheses, retrieve memory, and recommend action. By default it cannot change IAM, disable accounts, rotate credentials, alter firewalls, isolate endpoints, delete evidence, or deploy production code.
Security-sensitive improvements require evaluation and explicit approval where required. Supported adaptive detection is replayed before owner approval. Unapproved matches remain shadow evaluation; free-form model text is not executed as production security logic.
Practical deployment boundaries.
Human membership and role are resolved when a request is authorized. Service credentials are scoped by tenant, purpose, and expiry. Sensitive integration configuration is encrypted; notification delivery constrains destinations and preserves uncertainty when a provider's response is ambiguous.
No certification, independent audit, or guaranteed prevention rate is claimed here. Request deployment evidence and a review of your requirements before adoption. Report suspected vulnerabilities privately using the responsible-disclosure instructions below, without posting credentials or customer evidence in public issues.