When an AI agent or multi-agent workflow executes an unexpected high-impact action, traditional logs and raw prompt traces leave security teams guessing. Agent Black Box turns sanitized OpenTelemetry / OpenInference spans and MCP tool executions into a deterministic causal graph and evidence-backed incident timeline—without storing raw prompts, completions, or secrets.
127.0.0.1) via Docker Compose
Deterministic Forensic Architecture
Agent Black Box operates strictly downstream of your agent loop. It never sits in the execution blocking path.
Raw OTLP / OpenInference spans and MCP tool calls are stripped of prompts, completions, and bearer tokens before durable SQLite spooling.
Spans are mapped to canonical SecurityEvent v0.1 schemas with immutable SHA-256 hashes, source provenance, and trace continuity.
Rules ABB-001 through ABB-005 analyze input-to-action trajectories, authority drift, sensitive access, and verified state changes.
Events project into a Neo4j causal/authority graph linking documents, parent agents, tool executions, capability grants, and resources.
The Next.js Investigation Console provides deep-link run discovery, interactive timeline-to-graph highlighting, and provenance badges.
Engineered for Security & Platform Teams
Designed specifically to answer the hard questions after complex, delegated, or autonomous agent behaviors occur.
Connect agents using standard OTLP/JSON spans or the zero-dependency Python AgentRunTracer. All raw prompt bodies, completions, headers, and credentials are stripped before writing to disk; only structural metadata and SHA-256 hashes are preserved.
Full visibility into Model Context Protocol interactions. Inspect mcp_server_id, tool_id, authority grant IDs, declared capability scopes, and sensitive resource targets across multi-tool chains.
Evaluate canonical event sequences with rules ABB-001 (Untrusted Input to High-Privilege Action), ABB-002 (Intent-Action Drift), ABB-003 (Delegated Authority Escalation), ABB-004 (Sensitive Write), and ABB-005 (Corroborated State Change).
Explore the exact causal chain linking untrusted web/document inputs, prompt consumption, delegated child agents, service identities, and targeted database records in an interactive graph console.
Analyze multi-agent swarms for target convergence, knowledge transfer across agents, collective exploration, and outbound transfer policy violations—explicitly surfaced as derived analytical assessments.
Built upon OpenTelemetry, OpenInference, PostgreSQL, and Neo4j. Runs 100% locally on your workstation without cloud dependencies, outbound phone-home pings, or mandatory SaaS subscriptions.
Epistemic Integrity
Security teams cannot afford to confuse an agent's self-reported telemetry with independently verified system changes. Agent Black Box enforces this distinction across every tier of the product.
| Evidence Tier | Data Origin | Engine Status | Security Analyst Meaning |
|---|---|---|---|
| DIRECT (Unverified) | Agent / MCP OpenTelemetry & OpenInference spans | PROVISIONAL_UNVERIFIED · OPEN incident |
Indicates the agent self-reported executing a risky action; does not prove the target database or system state actually mutated. |
| CORROBORATED (Verified) | Independent target-system audit trail (e.g., PostgreSQL audit log) | ABB-005 verified · VERIFIED incident |
Confirms that an independent backend system verified the destructive state mutation with a cryptographic hash chain. |
| DERIVED (Analytical) | Graph Builder & Swarm Engine analytical projections | Read-only analytical qualification | Surfaces structural coordination and delegation patterns while preserving direct trace-level links to raw event IDs. |
Transparent Boundaries
127.0.0.1 via Docker Compose.agent_telemetry helper for agents and MCP tools.0.0.0.0.Privacy & Data Governance
The Agent Black Box Beta V1 software package runs 100% locally on your machine or private lab VM. It contains zero telemetry phone-home pings, zero analytics beacons, and zero external cloud tracking. Your agent traces, logs, and database records remain strictly on your device.
Information submitted via the waitlist intake form (name, email, role, and stack) is used exclusively to evaluate suitability for the invited closed beta and deliver your private download link and SHA-256 manifest. We never sell, rent, or share evaluator information with third parties.
Waitlist applicant records are retained only for the duration of the Beta V1 evaluation program. You may request immediate deletion of your submitted waitlist record at any time by contacting privacy@abbxai.com or security@abbxai.com.
Closed Beta Invitation
We are onboarding an initial cohort of AI engineers, MCP server developers, and application security practitioners to evaluate Agent Black Box Beta V1 locally and provide feedback on our detection rules and graph navigation.
To protect your privacy, this static page does not track you or transmit your data to third-party ad networks. You can submit directly via your email client or copy your structured request to clipboard.