What to log from runtime security decisions
Log every decision, allowed and blocked, with the action, the policy that fired and the outcome. Then send the records to your SIEM. HiddenLayer, Cisco and Bay document exports as of 16 Nov 2026, in webhook, Splunk, Sentinel, OpenTelemetry and OCSF formats.
Facts read on .
Which decisions should be logged?
Log all of them, not only blocks. A record of allowed actions shows what the control saw and lets you find a miss afterward, and a record of blocks shows what it stopped. Record the allow outcomes too, because a log of blocks alone cannot show a miss or measure a false positive rate later on.
Cisco says every decision in AI Defense is recorded. DefenseClaw answers allow, alert, ask or block for a prompt, shell command or tool call, so each of the four outcomes needs a field.
Sources: DefenseClaw guardrail docs.
What fields does a useful record hold?
Keep the time, the user or agent, the action type, the target, the policy that fired, the decision and the reason. Add a session identifier so a person can replay what led up to a block.
HiddenLayer documents recorded interactions and replayable sessions. Pillar says its activity view shows blocked decisions on session timelines.
Sources: HiddenLayer policy outcomes; Pillar Security agentic endpoint.
How do you send them to a SIEM?
Use the vendor's export in a format your SIEM reads, and confirm it carries the decision fields, not only alerts. OpenTelemetry defines a common log data model, and Cisco documents export to OpenTelemetry destinations.
HiddenLayer documents delivery of detect, redact or block outcomes to a webhook, Splunk or Azure Sentinel. Bay says its events reach the SIEM in OCSF format.
Sources: OpenTelemetry logs; HiddenLayer policy outcomes; Bay agent runtime enforcement; Bay AI agent investigation; Cisco: Protect enterprise coding agents with DefenseClaw and NVIDIA OpenShell.
What should you avoid keeping?
Avoid storing raw prompts and outputs longer than needed, since they hold the same sensitive data you are trying to protect. Cloudflare notes that its production mode logs only metadata.
Keep the decision and a reference to content held in a restricted store with its own retention rule.
Sources: Cloudflare AI Gateway Guardrails.
How do you use the logs?
Review blocked and allowed counts by action type each week, sample allowed actions near the policy edge, and feed repeat blocks into response playbooks.
Compare the log against the false positive budget you set, and keep the record for the audit period your regulator requires.
Who needs the logs?
Three groups read them: the security team for response, the application team for tuning and auditors for proof. Give each a view that fits and keep one source of truth.
For regulated work, the ability to show what was blocked and why is the purpose of the record, so test retrieval of a month-old decision.
What do vendors leave unclear?
Many vendor pages do not name the SIEM or SOAR connectors. The evidence files for Check Point, Palo Alto Networks, Noma and SentinelOne record named SIEM export as none found on 16 Nov 2026, though some have a logs API.
Ask for the field list and a sample export before you buy.
Sources: Check Point AI Security; Prisma AIRS API Intercept overview.
How long should records be kept?
Keep them as long as your audit and incident rules require, and no longer for raw content. Decision metadata can be held for a long period at low risk, while prompt and output text should follow a shorter rule.
Write the retention periods next to the field list so reviewers see both.
Sources
- DefenseClaw guardrail docs (cisco-ai-defense.github.io), read
- HiddenLayer policy outcomes (docs.hiddenlayer.ai), read
- Pillar Security agentic endpoint (pillar.security), read
- OpenTelemetry logs (opentelemetry.io), read
- Bay agent runtime enforcement (bay.io), read
- Cloudflare AI Gateway Guardrails (developers.cloudflare.com), read