Cybersecurity

The examples below describe potential use cases. Execution depends on available skills and connectors, customer permissions and qualification in the customer environment.

What the security team is asking for

An alert fires. The analyst needs the process tree from the endpoint tool, the authentication trail from the log platform, and what is actually on the server right now — and needs to be sure that looking does not change anything. Most of the shift goes into collecting, and the write-up is the part that gets cut.

• Follow one alert across the log platform, the endpoint tool and the machine itself.
• Audit a configuration across the whole estate and get a per-machine answer, not a sample.
• Prove what was checked, on which host, with which command — months later.
• Be certain an investigation cannot change a system by accident.

Why this vertical is the proof

Security is the hardest thing to build: real execution on real machines, reasoning across several systems, permissions that hold, evidence that survives an audit, and a human gate before anything is changed. It is built here first because it is the part that cannot be faked.

Hosts are the execution and evidence workers

A small agent runs on each enrolled machine — Linux or macOS. The model never touches the machine; it asks the host worker to run an approved diagnostic, and the worker returns the command, the exit code and the output.

A host joins the fleet with one install command and an enrollment token; revoking the host unpairs it.

Read-only execution policy is the default on every enrolled host; the full policy is an explicit, per-host choice.

Host investigations show the commands that ran, their exit codes and output, per host.

The skills an investigation may use

Skills ship with the release, come from AcTree's signed skill catalogue, or are your own uploads, and each workspace enables or disables each one.

The built-in catalogue is written for investigation rather than for chat — each skill names the evidence to collect, in what order, and what it must not touch:

• linux-host-triage
• windows-host-triage
• active-directory-investigation
• identity-dns-edr
• file-provenance
• infrastructure-incident-rca
• exchange-server-triage
• network-firewall-triage
• storage-filesystem-triage
• kubernetes-cluster-triage
• virtualization-platform-triage

Each skill is a versioned package. Uploads are checked before they install, every version is archived, and each workspace enables or disables each skill.

How an investigation is carried out

An agent is a model, a set of approved skills, tools and a policy, applied to a scope.

Agents reach your systems through the MCP servers you connect in the console. A token you enter there is sealed, never shown again, and acts only with the permissions your system gives it; a connector that runs on one of your hosts keeps its token in that host's own connector file.

Execution follows the configured host policy. A command can be approved, denied or referred for human review; individual approval depends on that policy.

Where it runs

Customer-hosted deployments use packaged deployment configuration. An offline deployment requires an accepted bundle, local model and disconnected installation tests.

Bring your own model — OpenAI-compatible endpoints, Ollama, Anthropic, or Cloudflare AI Gateway / Workers AI — with keys envelope-encrypted at rest.

Related guides and agents

See how SOC teams use private AI agents for investigations and triage:

• AI for SOC analysts — how security teams use private AI agents to investigate alerts and audit controls.
• Security agent — the agent built for SOC analysts, connecting to log platforms, endpoint tools, and servers.

Customer-hosted installations support an identity provider through OIDC with PKCE. Hosted workspace SSO must be confirmed for the agreed deployment.

Customer-controlled inference and execution boundaries.