Frequently asked questions.

Aegis is a real-time execution firewall for LLM applications and autonomous AI agents. It runs deterministic guardrails on prompts and responses, while actively inspecting every tool and MCP call an agent makes — authorizing or blocking actions before execution.

Integration is drop-in. Aegis provides an OpenAI-compatible API gateway. You simply update your existing client's base URL and API key. No changes to prompt templates, agent logic, or streaming handlers are required.

Yes — this is our core differentiator. Aegis inspects JSON-RPC MCP messages and agent tool calls against per-tenant allow/deny policies, least-privilege schemas, and dangerous-action boundaries before tools run.

Aegis defends against prompt injection, indirect (data-borne) injection, jailbreaks, data exfiltration, destructive shell/SQL commands, unauthorized funds transfers, and tool output poisoning.

No. The guardrail engine runs pattern-based deterministic checks in-process, without round-trips to secondary LLMs. The prompt and response rails add 2.9 ms at p95. A turn that also carries tool or MCP calls goes through the full execution rail, measured at 14 ms at p95 against a 50 ms budget. Both figures are measured on our benchmark suite, not modelled.

Aegis inspects streaming tokens and buffers streaming tool calls in real time. Tool arguments are evaluated against policy before execution begins, closing the bypass vulnerability common in naive proxies.

Every database record is scoped to its tenant and queried through a centralized data access layer that enforces tenant boundaries on every read and write. There is no cross-tenant read vector.

All control-plane actions and policy modifications are written to an append-only log where each entry embeds the SHA-256 hash of the preceding record. Editing or deleting an entry is refused outright — by database triggers, and on PostgreSQL by a privilege boundary as well — and would break the chain, which a verification pass reports along with the first entry it cannot account for. Appending is the one write the log permits, so an entry added directly to the database still moves the chain head; comparing that head against a recorded baseline is what catches it, and it is a standing part of our backup and restore checks.

Yes. Aegis streams events natively as ArcSight CEF over syslog (UDP/TCP) or NDJSON over HTTPS, enabling instant ingestion into Splunk, Microsoft Sentinel, Datadog, or Elastic.

Aegis automatically generates reports mapping runtime verdicts to the OWASP LLM Top 10 (2025), high-risk obligations under the EU AI Act, NIST AI RMF 1.0 subcategories and MITRE ATLAS techniques — each with an explicit coverage verdict, including the controls a runtime gateway does not address, and hash-chain integrity verification metadata.

Aegis deploys directly inside your private cloud or on-premise infrastructure via Kubernetes (Helm chart) or Docker Compose. Your sensitive prompts and data never leave your security perimeter.

Yes. Policies are scoped per tenant and environment (development, staging, production) with immutable versioning, atomic rollback, and role-based access controls for security administrators and developers.