The Arcus Security Gateway is provided to manage and control API traffic. You agree not to use the platform for malicious activities, including but not limited to DDoS attacks, unauthorized scanning, or distributing malware. Violations will result in immediate workspace suspension.
To ensure stability across all tenants, Arcus enforces fair use rate limits on API requests traversing the gateway. Sustained bursts exceeding your workspace limits may result in temporary throttling (HTTP 429). Enterprise plans are available for higher throughput requirements.
We strive for high availability and minimal latency, but Arcus is provided "as is" without warranty of any kind. We are not liable for any lost data, lost revenue, or downtime resulting from platform usage.
Arcus is a control plane for traffic dispatched by your own AI agents. You are responsible for every agent connected to your workspace: if one causes harm at a system you have no authorisation to reach, that is a breach of this agreement by you. Dispatch is deny-by-default — a message goes out only where a policy in your workspace explicitly allows it, and neither we nor a policy you have configured may be used to grant an agent access to a third party's internal systems beyond what you are yourself entitled to reach.
Keys are issued to your workspace alone, and the raw value is shown exactly once at creation — it is never retrievable afterwards. Treat one as you would a password: do not commit it to source control, log it, or place it in an agent prompt. Where you bind a key to a specific agent identity, requests presenting it under any other agent are refused. You are responsible for revoking any key you believe is exposed; we cannot recover a leaked credential on your behalf.
Every dispatch is evaluated against your policies and scanned for sensitive data before it leaves your perimeter. That requires us to read the body of each message: content is inspected, matched against the rules you configured, and recorded in your transfer log. Live credentials found in an outbound payload are refused on every plan. Your audit trail is a hash chain, so records can be read and verified but not edited or deleted from your side. Retention of those records is bounded by your plan and is disclosed in the workspace.
Per-agent limits exist so one workspace cannot degrade another: exceeding 60 dispatches per agent in a 60-second window quarantines that agent for 300 seconds, and repeated triggers may draw a longer operator-applied throttle. Refused requests return HTTP 429 with a Retry-After value. Key and guardrail-rule ceilings are set by your plan and shown in the workspace. We may suspend a workspace that breaches section 1, or that uses the gateway to attack the platform itself, and suspension carries no refund for the remainder of the billing period.