Industry Solutions

How do you build AI agents for cybersecurity?

Cybersecurity agents enrich alerts, summarise indicators, draft detection rules and assemble incident timelines. They should operate with least-privilege, read-only tools and never auto-remediate; a human analyst decides and acts. Plugsky provides an OpenAI-compatible API with function calling, 30+ models and embeddings to build that bounded, fully logged loop.

Key facts

Agent stackChat completions, function calling and embeddings are live
Use casesAlert enrichment, IOC summaries, detection rule drafts, incident timelines
Least privilegeRead-only tools scoped per integration; no auto-remediation
AuditLog prompts, tool calls, sources and model versions for every run
Models30+ models; small aliases for classification, stronger ones for synthesis
DeploymentPlugsky cloud, your VPC, on-prem and air-gapped options
Pricing modelFlat monthly self-serve plans with unlimited fair-use usage
RoadmapBatch and assistants endpoints are coming soon

TL;DR

  • Enrich and prioritise alerts for analysts instead of closing them.
  • Summarise indicators and attacker behaviour with cited sources.
  • Draft detection rules and incident timelines for human review.
  • Never allow autonomous response actions such as blocking or isolation.
  • Log every tool call so a security incident can be reconstructed later.

How it works, step by step

  1. Choose one workflow such as alert enrichment or incident timeline assembly.
  2. Expose least-privilege, read-only tools for logs, threat intelligence and case systems.
  3. Ground summaries in retrieved evidence and require source references per claim.
  4. Draft detection rules and timelines as proposals for analyst review.
  5. Forbid automated response actions and enforce that in code and network policy.
  6. Log tool calls and model versions, then measure triage time and false-negative rate.
1Choose one workflowsuch as alertenrichment or2Exposeleast-privilege,read-only tools for3Ground summaries inretrieved evidenceand require source4Draft detectionrules and timelinesas proposals for5Forbid automatedresponse actionsand enforce that in6Log tool calls andmodel versions,then measure triage

Try it yourself

Open the AI API key security checklist →

Security workflows suited to agents

Security operations generate more alerts than analysts can read, which is where agents help most:

  • Alert enrichment: attach asset context, user history and prior incidents to each alert and rank it.
  • IOC summaries: consolidate indicators from multiple sources into a short, cited brief.
  • Detection rule drafting: propose rules or queries for an analyst to validate and test.
  • Incident timelines: assemble events from logs and notes into a reviewable chronology.

Analysis and response decisions remain with the security team.

Architecture with least privilege

Treat the agent as an untrusted insider and design accordingly:

  • Scoped tools: read-only access to logs, intel and case data, with per-integration credentials.
  • No response tools: no blocking, isolation, quarantine or credential reset calls.
  • Retriever: embeddings over runbooks, prior incidents and detection content.
  • Composer: outputs with citations and explicit uncertainty.
  • Analyst gate: a human reviews, decides and performs any action.

Rotate tool credentials, cap query scope, and log every call for later forensic review.

Data handling and isolation

Security telemetry can contain personal data, credentials and sensitive infrastructure detail. Redact or reference rather than copying secrets and raw user records into prompts, set retention deliberately, and choose a processing region that satisfies your obligations. Use scoped keys, RBAC, SSO and audit logs across the platform. Where the environment requires it, deploy in a VPC, on-prem or air-gapped setting with the same compatible API, so detections, prompts and evaluations stay portable. Keep the agent out of the critical path: if it is unavailable, alerting and response must continue unchanged.

Pilot and evaluation

Prototype on plugsky-micro or plugsky-lite with historical alerts and a runbook corpus, then use the 14-day full-access trial to test a stronger model on enrichment and timeline building. Measure triage time per alert, ranking quality, citation support and false-negative rate on known incidents. Self-serve plans are flat monthly with fair-use usage, so alert storms do not create variable bills. Roll out to one queue, keep prompts versioned, and re-test whenever data sources or detection content change.

Honest comparison

CapabilityPlugskyTypical agent frameworkBuilding in-house
Tool callingLive function calling with per-integration read-only credentialsFramework-specific adaptersYou build and maintain
Response safetyNo auto-remediation; you enforce read-only by designVariesYou design controls
AuditLog prompts, tool calls and model versionsPartial or add-onYou build all of it
DeploymentCloud, VPC, on-prem and air-gappedUsually cloud-onlyYou operate GPUs and serving
PricingFlat monthly self-serve, fair-use usageVaries by framework and modelGPU plus operations cost

Frequently asked questions

Can the agent block or isolate a host?

No. Forbid response actions and keep tools read-only; an analyst decides and performs any remediation through your existing tooling.

Which endpoints do security agents use?

Chat completions, function calling, JSON mode and embeddings are live. Batch and assistants endpoints are coming soon.

How do we keep credentials safe?

Use scoped, rotatable credentials per integration, redact secrets from prompts and logs, and audit every tool call.

Is there a free plan?

Yes. The free plan includes plugsky-micro and plugsky-lite with no credit card, enough to prototype alert enrichment and summarisation.

Can we run it air-gapped?

Yes. Air-gapped deployment is available, and the compatible API means the agent code and evaluations run unchanged.

How does pricing work?

Self-serve plans are flat monthly with unlimited fair-use usage. See the live pricing page for current plans and enterprise options.

How should we measure it?

Track triage time per alert, ranking quality, citation support and false-negative rate against known incidents.