Agentic Capabilities
What is the Mitigant MCP server?
The Mitigant MCP server connects AI agents to the Mitigant platform through the Model Context Protocol (MCP), the open standard agents use to work with external tools. It turns AI agents into reliable offensive security operators. Agents can plan a security exercise, carry it out safely, read the evidence, judge which security controls held, clean up and report back. It starts with Cloud Attack Emulation and will cover more of the Mitigant platform over time.
Why give AI agents offensive security capability?
Testing security controls is a job, not a single action. It means planning, choosing the right techniques, executing safely, reading logs, judging results and writing up findings. Many teams have no dedicated red/purple team to do that work. The MCP server lets AI agents take on that job with the safety and reliability an enterprise environment needs.
Which AI agents and MCP clients does it work with?
Any MCP client works. Setup is documented for Claude Code and Kiro. Other clients such as Cursor, Codex, VS Code or ollmcp connect the same way: generate an API key and point the client at the Mitigant MCP endpoint with the key in the request header.
How do AI agents connect to Mitigant?
Generate an API key in the Mitigant dashboard, then add the endpoint https://mcp.mitigant.io/mcp to the agent with the key in the apiToken header. Ready-to-copy setup for Claude Code and Kiro is on the Mitigant MCP page.
What do security teams use it for?
- Detection validation: Agents validate whether threat detection mechanisms work end to end, from log ingestion and SIEM configuration to detection rules and alerting. For example, an agent can check whether any detection covers a given attack technique.
- Red/purple team exercises: Agents plan and carry out red/purple team exercises, from a single technique to a full adversary emulation, and report how defenses held up. For example, an agent can emulate a known threat actor's techniques against a chosen cloud account.
- Security control validation: Agents validate whether preventive, detective and compliance controls work as intended, and turn findings into evidence of real, exploitable risk. Every run leaves a reproducible record for audits. For example, an agent can take a findings report and test which findings an attacker could actually exploit.
- Incident response readiness: Agents run realistic breach scenarios so teams can test whether people, processes and technologies respond as planned, without the usual preparation overhead. For example, an agent can run a multi-step attack scenario during a tabletop exercise and document how the response unfolded.
Is it safe to let an AI agent carry out cloud attacks?
Yes. Agents reason and plan freely, but every technique they execute comes from a curated catalog of tested attacks mapped to MITRE ATT&CK and MITRE ATLAS, not code the agent writes. Execution needs approval, stays within the chosen scope and cleans up automatically, so the environment is left as it was.
Do teams need offensive security expertise?
No. A plain-language request is enough, for example "check whether an attacker disabling S3 logging would be detected". The offensive security expertise is built into Mitigant; the agent handles the planning and the reporting.
How is this different from asking an LLM to write attack scripts?
An LLM can write attack code, but it cannot execute it safely, prove what the controls detected or undo the changes. Mitigant's techniques are tested, MITRE-mapped and reversible, run only with approval and within scope, and clean up automatically. Teams get evidence they can trust instead of untested scripts to vet and clean up by hand.
What are AI-powered attack recommendations?
Mitigant recommends the attacks that matter most for a specific cloud account. It analyzes the account for weaknesses and misconfigurations, then recommends the attacks with the greatest impact on that setup. Each recommendation names the attack, the target cloud service, the potential impact and the MITRE ATT&CK tactics involved. Teams test high-impact weaknesses first and prove a weakness is exploitable before investing in remediation. Recommendations appear in the attack library.
What is AI-powered attack report summarization?
Every attack emulation gets an AI summary automatically. It states what happened, which services and resources were targeted, the impact if the attack had been real, and prioritized remediation steps, with MITRE ATT&CK and MITRE ATLAS mappings. Security teams and leaders get a clear picture without reading pages of technical logs. The summary sits in the AI Summary tab of each report.
Learn more: Feature Release: Leveraging AI for Adversary Emulation