Firefish docs

Architecture

Firefish is a gateway boundary with scanners, policy, runtime mediation, tool governance, audit, and operator review.

Core components

Prompt/content scanner

Normalizes text, classifies source type, applies deterministic rules, local lexical signals, optional anomaly checks, and configured judge routing.

Policy engine

Maps risk score, attack type, and source context into ALLOW, FLAG, REDACT, SANDBOX, REQUIRE_APPROVAL, or BLOCK.

Runtime broker

Mediates model/runtime calls so protected chat traffic can be evaluated before and after model interaction.

Output scanner

Checks model output for secrets, unsafe content, and policy violations before the response reaches a user.

Streaming gate

Inspects streaming response chunks while output is still moving through the gateway.

Tool validator

Evaluates proposed tool calls against action class, destination risk, secret exposure, reversibility, and user goal alignment.

Tool result sanitizer

Redacts sensitive material from tool results before model or operator consumption.

Audit logger

Stores decisions, reason codes, redacted previews, traces, and latency metadata for review.

Dashboard

Shows mission control, threat lab, tool firewall, benchmark center, demo flows, and technical proof points under /app/.

Security pipeline

Classify

Normalize input and identify source type.

Detect

Run deterministic and configured analysis layers.

Decide

Apply policy without lowering global block thresholds.

Audit

Save redacted traces for operators and reviewers.