Skip to content

AI model

AI SIEM uses an Amazon Bedrock model (Anthropic Claude) to write incident response playbooks and the qualitative lead in reports. Every deployment ships with a sensible default, so this works out of the box with nothing to configure. Detection itself is deterministic — rules and ML baselines decide what is a threat; the model only phrases the response. Identifiers are de-identified before anything is sent to Bedrock (see the FAQ).

Default model and Bedrock access

The default model is Claude Haiku 5.5 (anthropic.claude-haiku-5-5). AI SIEM adds the regional inference-profile prefix (us., eu. or apac.) for your deployment region automatically, so you only ever deal with the base id.

Bedrock model access is granted per model, per account. Before AI features work, a user with AWS Marketplace permissions must enable Claude Haiku 5.5 once in the Bedrock console (Model catalog → open the model → request access), and submit the Anthropic use case form if your account has not done so before. Until then, AI playbooks fall back to the built-in template and the FAQ's Bedrock row reports access denied.

Upgrading from 26h4c or earlier? The previous default was Claude Haiku 4.5, which is past its AWS end-of-life floor (2026-10-15). After the upgrade your deployment calls Haiku 5.5, so enable Haiku 5.5 access before or right after upgrading — access to 4.5 does not carry over. Then open Settings → AI and click Test this model to confirm. If an admin had already chosen a model in Settings → AI, that choice is kept and the new default does not apply.

Choosing a model

An admin can change the model under Settings → AI in the editor:

  1. Pick a model from the recommended list. Each entry shows its relative cost (lower / standard / higher) so you can weigh cost against capability.
  2. Click Test this model to run one real playbook generation and preview the result — confirming the model works in your account before you commit.
  3. Click Apply to make it live. The change takes effect across the deployment within a minute, with no stack update.
  4. Revert to deploy default at any time to go back to the model your stack was deployed with.

Advanced users can enter an unlisted model id, but a listed model is recommended — those are the models validated to work with AI SIEM's prompts.

Cost

The model choice affects only the per-playbook AI cost, and playbooks are cached and generated on demand, so a higher-cost model does not cause runaway spend. The picker shows a relative cost band rather than a dollar figure; see current Amazon Bedrock pricing for rates, and pricing for the overall cost picture.

Model retirement

AWS periodically retires older model versions. If the model you are using nears its retirement date, the editor's FAQ panel shows an amber advance warning prompting you to choose a successor in Settings → AI. AI keeps working until you switch, so there is no sudden outage — and switching needs no redeploy.

If a model ever becomes unavailable before you switch, AI SIEM falls back to a built-in deterministic playbook and tells you why, so incident response is never left empty.