An offensive engine and an audit toolkit.

A penetration-testing platform that runs as a coordinated multi-agent system. Specialist agents for reconnaissance, exploitation and reporting hand work to each other through native tool calling, and untrusted target output is treated as hostile input rather than as context. Findings leave the run as structured, schema-validated records.

• Pipeline · specialist recon, exploit and reporting agents with native tool calling
• Injection · untrusted target output screened before it reaches agent context
• Runtime · cost tracking and prompt caching across the whole run
• Output · structured, schema-validated findings

pipx install cyberai·Apache-2.0·Python 3.11+

Where CyberAI attacks, MST inspects. A scan opens a connection, negotiates a protocol version, calls what the server exposes and reads what comes back — a finding exists because a probe observed a behaviour on the wire, not because a classifier scored a description. Nothing leaves the host, output is SARIF, and a probe that could not run is reported rather than silently dropped.

• MCP · STDIO and streamable HTTP on both spec routes, nine probe families
• A2A · AgentCard audit, card poisoning, routing hijack, delegation-mesh escalation
• Brokers · MQTT and AMQP posture, retained-payload injection confirmed by read-back
• ABFP · five-dimension fingerprinting with baseline diffing for drift and rogue agents

Detection covers the full OWASP Agentic Top 10:

1. ASI01 Agent Goal Hijack
2. ASI02 Tool Misuse & Exploitation
3. ASI03 Agent Identity & Privilege Abuse
4. ASI04 Agentic Supply Chain Compromise
5. ASI05 Unexpected Code Execution
6. ASI06 Memory & Context Poisoning
7. ASI07 Insecure Inter-Agent Communication
8. ASI08 Cascading Agent Failures
9. ASI09 Human-Agent Trust Exploitation
10. ASI10 Rogue Agents

pipx install mas-sentry-toolkit·AGPL-3.0-or-later·Python 3.11+