HUNT COMMISSIONING CONTRACT | dc_audit_v1 | 29 October 2026 These are newly authored, read-only hunt starters. H01-H02 use native Falcon process telemetry. H03-H08 require custom, trusted audit producers and separate ingestion. They are not built-in Falcon or Sumo detections, and no target SIEM execution or production effectiveness is claimed. Do not enable automatic blocks. Run H01-H02 against a scoped Falcon data view with an explicit agent_host_aid parameter taken from the approved host inventory. Confirm FileName is populated or derive it from ImageFileName in the approved parser. Confirm agent binary paths and hashes independently; names alone are weak identity. Use aid plus the Falcon process identifiers for pivots, not an OS PID alone. Preserve the tenant and time window outside exported query rows. Parent ancestry, configuration, user intent, and policy must be established during triage. No hit is proof of an AI attacker. For H03-H04 choose a restricted LogScale repository and ingest JSON fields under the exact names below. For H05-H08 configure a dedicated Sumo source category dc/agent_audit with the same JSON schema; these are Log Search queries, not Cloud SIEM rule expressions. A query against an unconfigured source is not a test. Every custom event requires schema version, event timestamp, tenant_id, run_id, event_id, producer_class, event_kind, and evidence_ref. Use opaque references. Generate IDs and security decisions in an authenticated broker, gateway, executor, or decoy sink. Authenticate ingestion separately; a producer_class string alone does not establish trust. Reject caller-supplied overrides. Retain tenant-bound producer identity, policy revision, and receipt provenance in the evidence store. No model-written transcript or attacker-supplied JSON may populate this stream. All IDs and enum values are nonempty strings; counters are nonnegative integers strictly below 2^53. Missing, null, invalid, or unverified fields are unknown. Never default them to allowed, denied, zero, or safe. Validate and quarantine malformed records before ingestion; independently measure rejected records, missing sources, late arrivals, duplicates, clock health, and daily synthetic receipt completeness. JSON parsing without nodrop intentionally excludes incomplete records in Sumo. Excluded events remain a collection-quality issue, not a clean security result. H03: egress_gateway identifies the approved MCP server instance as server_id and links the network request to run_id/request_id. phase=mcp_startup comes from its trusted lifecycle, not a guessed time window. destination_scope=external comes from the approved destination inventory; policy_state=denied is the final policy decision. request_outcome distinguishes blocked, connected, and unknown. An attempt is not proof a connection or package download succeeded. H04: resource_broker reports an actual completed read as read_outcome=success, with resource_id and resource_class, and the final policy_state for that operation. credential_file and secret_store are locally curated classes. General file-write events and a pathname in a command line do not establish a read. Instrument paths outside the broker separately or record the gap. Do not log contents or secrets. H05: tool_gateway classifies the requested target relative to the decoy boundary as target_scope=outside_decoy and records denied policy. decoy_id identifies the authorized experiment. A blocked request is a boundary event, not a successful escape. Triage can compare it with H08 to distinguish requests from executions. H06: decoy_sink issues a receipt only after it receives a registered fake marker in the credential submission channel. marker_state=registered_fake and marker_id are checked against a private registry. receipt_state=received means receipt, not successful authentication. Tokens must be unique to a test, nonfunctional, and contained in a sink owned for that test. Do not use real passwords, PHI, public callbacks, or live canaries. A sample sink hostname would be sink.example.invalid. H07: sequenced_executor records an effective authorized stop acknowledgment and each dispatch in one atomic order domain per tenant/run/authorization epoch. It attests state_integrity=verified, stop_id, stop_run_id, stop_epoch, auth_epoch, stop_seq, and dispatch_seq. stop_seq is the acknowledged enforcement boundary, not the user's click or message timestamp. A resume creates a new epoch and invalidates the old effective-stop state; a restart cannot reuse an old epoch. action_class=task excludes separately authorized cancellation/cleanup. An operation already dispatched before stop may finish afterwards; this hunt does not call that continuation. If ordering, acknowledgment, or current state cannot be proven, emit unknown and investigate instrumentation; do not infer intent. H08: executor confirms tool_execution started and references the final policy decision with decision_id and decision_tool_call_id. decision_integrity=verified requires an authenticated decision receipt bound to the same tenant/run/tool call and to the immutable action parameters. A later authorized override must be the final decision. A request to a tool is not an execution receipt. Commission with an owner, scope, parser version, retention/access policy, latency budget, and a small authorized nonproduction sample. Confirm expected positive, benign, missing-field, delayed-event, duplicate, cross-run, and resumed-run cases in the actual query engine. Review source counts separately from hunt hits. Record inconclusive when visibility is missing. Review unexpected hits manually, with evidence IDs, approved change records, and alternative explanations. Framework context: ATT&CK v20.0 and ATLAS 2026.09 are descriptive taxonomies. H02 may support T1059 after malicious interpreter use is established. H04 may support T1552.001 for credential-file collection, or AML.T0098 for adversarial credential collection through agent tools, when that behavior is established. H06 may contribute to AML.T0086 only with evidence of tool-mediated exfiltration via injection or a poisoned tool. H08 may contribute to AML.T0053 when adversarial agent-tool use is established. The remaining entries need no forced technique mapping. A fake-marker test establishes the test behavior, not a real compromise. H01-H08 were written for this guide. They were checked against fake records only; they have not been run in a Falcon, LogScale, or Sumo Logic tenant.