Skip to content
Cisco AI Defense logo
CiscoAI Security

Glossary — DefenseClaw

Overview

When two words could mean the same thing, this page says which one we use and what it means.

Terms

Action — The enforcement outcome of a verdict. One of allow, warn, block, quarantine. Distinct from severity.

Admission — The install-time gate that decides whether a skill/MCP/plugin/tool is allowed into the agent runtime. Implemented by admission.rego.

Admission override — An operator decision to install an artifact despite a block verdict, recorded in the audit store with a required reason.

AIBOM — AI Bill of Materials. CycloneDX-shaped inventory of models, datasets, skills, and plugins installed on the host.

Audit bridge — The translator that maps persisted audit events into gatewaylog.Event lifecycle rows for gateway.jsonl.

Audit store — The local SQLite database of every decision and verdict. Authoritative; all other sinks are copies.

Bifrost — The provider-routing layer used by the Go gateway to call configured LLM providers.

ClawShield — The binary/archive/image scanner.

CodeGuard — The generated-code scanner.

Correlation ID — A string carried end-to-end across a single agent turn so every emitted event can be joined back into one story. Format: req_01HX... (ULID-like).

Directionprompt | completion | tool_call. The role of the content being inspected.

Drift — Content of an already-approved artifact changed since approval. Detected by snapshot-diff in the watcher.

Enforcement — The physical-layer action that follows an admission decision (quarantine, restore, skip). Distinct from admission which is the decision.

Event — A structured record of something happening. Produced by every subsystem; persisted in the audit store; forwarded to sinks.

Event filter — The sink-level min_severity and actions filters that decide whether an audit event is delivered.

Finding — A single detection emitted by a scanner or rule. Identified by a prefixed string (secret:aws-key, injection:shell, pii:email). Multiple findings compose a verdict.

Guardrail — The LLM traffic inspector: the proxy + scanner pipeline that inspects prompts, completions, tool calls.

Judge — The LLM-based adjudicator used in regex_judge and judge_first detection strategies.

LLM judge client — The Go client under internal/gateway/llm_judge.go that calls the judge, enforces timeouts, and caches verdicts.

Master key — The default API bearer token minted at install time.

MCP — Model Context Protocol. The standard for agent tool providers.

Modeobserve | action | disabled. Controls whether the guardrail logs or enforces.

Notification queue — The in-process queue used by the guardrail path to inject pending enforcement notifications into model-facing system messages.

OpenShell — The CLI that invokes shell subprocesses on behalf of agents. Sandbox-wrapped by DefenseClaw.

Pack version — A hash of the compiled rule pack + suppressions snapshot. Used to version-scope the verdict cache.

Policy snapshot — An immutable compile of Rego modules + YAML rules + data.json overlays. Swapped atomically on reload.

Profile — A named rule-pack directory: default, strict, permissive, or custom.

Quarantine — The physical move of a blocked artifact to ~/.defenseclaw/quarantine/.

Redaction — Scrubbing sensitive bytes from events before they leave the gateway.

Scanner — Any component that produces a ScanVerdict.

Scanner modelocal | remote | both. Controls where the scanner runs.

Sensitive tool — An MCP tool marked in sensitive-tools.yaml for tighter result inspection.

Sidecar — The Go process (defenseclaw-gateway) that owns policy, audit, and the guardrail proxy.

Sink — An audit_sinks[] destination for audit events, currently splunk_hec, otlp_logs, or http_jsonl.

SSRF — Server-Side Request Forgery. The firewall's SSRF guard mitigates this.

Strategyregex_only | regex_judge | judge_first. Controls how the guardrail decides.

Suppression — A declarative rule that drops or strips content to eliminate false positives.

TUI — The terminal UI (defenseclaw tui) that surfaces everything the CLI can show, interactively.

Verdict — The structured output of a scan. severity, findings, action, reason, confidence.

Verdict cache — In-process cache for guardrail verdicts, keyed by content/direction/model/kind in the current source.

Watcher — The filesystem observer that triggers scanners on install/change.

Related