Cybersecurity Products

Security vendors build assistants that read alerts, explain detections and draft code. The product itself handles hostile input every day. A starter for this work has to assume that some prompts arrive from the very traffic the product exists to study.

The chain opens with injection screening, then masks API key and bearer token shapes in the prompt. Two phrases end a request outright: private key and production credential. A responder who pastes a live secret into a support ticket gets a refusal rather than a model call.

Generated code passes a sanitizer that is set to refuse on a match rather than warn. That protects a build pipeline from an unsafe snippet nobody read closely. The check is pattern based, so an unfamiliar unsafe construct can still get through. Treat it as one layer under your normal secure development review.

Requested tool calls are examined on the gateway, and no allowed or denied tool list ships here. Naming the tools an agent may call stays your decision. Output scoring holds replies to a combined 0.82, relevance 0.78 and accuracy 0.80, with a floor of 120 characters and two sentences.

The review point carries no configuration, so it escalates nothing until your team enables it. Audit marking flags an evaluated request and never stops one. Control evidence here speaks to SOC 2 Type II, ISO 27001 and NIST CSF 2.0 questions about secure development. Vulnerability handling, coordinated disclosure and change control remain outside the starter.