AIForSRE LABS · PREVIEW

Practice decisions, not definitions.

Fictional production incidents. Synthetic data. Nothing leaves your browser.

RAG SECURITY

Poisoned Runbook

A newly ingested runbook is dominating retrieval and steering an operations agent toward an unsafe action.

  • Retrieval source changed 30 min ago
  • Same chunk appears in 82% of answers
  • Agent write-tool requests increased 5×
Recommended response

Quarantine the source, disable write tools, preserve traces and rebuild the affected index after review.

MCP SECURITY

MCP Over-Permission

A read-oriented support tool can access financial and credential tables through its backend identity.

  • Tool description says customer lookup
  • Database role is schema owner
  • Sensitive tables visible in catalog
Recommended response

Replace the backend identity with least privilege, split tool capabilities and audit historical access.

AI SRE

Runaway Agent Loop

An agent repeatedly calls the same troubleshooting tool and token spend is climbing.

  • Tool call count > 60/session
  • No terminal condition reached
  • Cost/request up 14×
Recommended response

Stop the loop, enforce iteration/tool budgets, add terminal conditions and inspect the trigger.

RELIABILITY

Provider Outage

Primary model provider latency exceeds 20 seconds and error rate is rising.

  • Provider 5xx rising
  • Queue depth 900
  • Fallback never tested
Recommended response

Activate tested degraded/fallback mode, protect queues and compare behavior before broad failover.

RAG AUTHORIZATION

Cross-Tenant Retrieval

A user receives a citation from another business unit's restricted document collection.

  • Vector similarity is correct
  • ACL metadata missing on one ingestion path
  • No source-system authorization check
Recommended response

Disable affected retrieval path, restore ACL metadata, re-index and test tenant isolation.

AI COE

CoE Intake Decision

A business unit wants to launch a customer-facing agent in two weeks using sensitive CRM data and write-capable tools.

  • No risk tier assigned
  • No SLO/on-call owner
  • Provider not in approved catalog
Recommended response

Classify risk, route through architecture/security/production gates, and narrow capabilities for the first release.

PRIVACY

No uploads. No production data.

Future interactive labs will keep answers and progress in local browser storage unless a user explicitly opts into an account later.