Inputs & Documents
Stops poisoned prompts, RAG chunks, and document instructions before they become trusted context.
RAG / Document Firewall, prompt scanning, source-aware decisions.
Local-first AI agent firewall
Stop AI agents from trusting poisoned content, leaking secrets, or taking unsafe actions.
Purple Firefish is a local-first AI agent firewall for LLM apps, RAG systems, and tool-using agents. It scans documents, prompts, tool calls, model outputs, and sensitive data before untrusted instructions become trusted behavior.
Local-first. No raw prompt logging by default. Synthetic benchmarks. Audit-ready evidence.
Threat path
Where Firefish interrupts unsafe behavior
Poisoned document
Retrieved context
Agent proposes tool call
Firefish evaluates
Block / approve / redact
Audit evidence
Run local benchmark to populate proof metrics.
No benchmark numbers are shown until data/benchmark_results/latest.json exists.
Three protection lanes
The homepage version of Firefish
Secrets & Outputs
Detects sensitive values before they leak through prompts, tool arguments, or model responses.
DLP, secrets detection, tokenization, output redaction.Agent Actions
Evaluates proposed tool calls before writes, sends, exports, deletes, or code execution can happen.
Agent Capability Firewall, approvals, dry-run enforcement.
Local and private AI
Designed for private deployment
Run Firefish self-hosted with local/private model stacks. Raw logging stays disabled by default, and evidence is stored as audit-safe metadata.
Operator view
Evidence without noise
Security Ops pulls together blocked actions, redactions, quarantined documents, approvals, incident timelines, and CI regression signals.
Honest limits
Firefish does not guarantee complete protection or replace authorization and human review. It is one local-first layer in a broader AI security program. Read the limits.