1200kmSECURITY RESEARCH
Loading interactive filters…

1200KM / simulation

T1700 AI-Enabled Exfiltration — Attack Simulation

Adversaries may utilize Direct Prompt Injection, Indirect Prompt Injection or other methods to instruct enterprise AI agents to exfiltrate data. When enterprise systems incorporate the use of AI assistants or other AI agents, adversaries may abuse these systems to weaponize sanctioned tools available to the agent. As an AI agent processes queries or ingests data, an injected prompt can cause the agent to use its tool connections in unauthorized…

Complete technique description

Adversaries may utilize Direct Prompt Injection, Indirect Prompt Injection or other methods to instruct enterprise AI agents to exfiltrate data.[Prompt-Injection Exfil][Anthropic Report Sep 2026] When enterprise systems incorporate the use of AI assistants or other AI agents, adversaries may abuse these systems to weaponize sanctioned tools available to the agent.[Google Security State of Prompt Injections]

As an AI agent processes queries or ingests data, an injected prompt can cause the agent to use its tool connections in unauthorized or unintended ways. When AI agents are induced to exfiltrate data, other exfiltration techniques likely apply as well to transfer the information out of the network, such as Exfiltration to Code Repository or Exfiltration Over Web Service.[Google Cloud AI Risk Sep 2026][Barracuda Prompt Injection Phishing Emails]

Official technique definition · Open in the interactive matrix

Detection telemetry and configuration

These contracts are selected through exact technique-to-data-component relationships or explicitly labelled planning overrides. Configure a collector at the relevant observation point; a telemetry name alone is not coverage.

Application Log Content · source-mapped

DC0038 collection page — Select the application event family required by the detection and document its native schema before normalization.

Providers and collection configuration

  • Enable the application audit category or instrument the owned lab application at the authorization/action boundary.
  • Define a schema with timestamp, service, actor, object, action, result and correlation ID. Export structured records through an authenticated collector.
  • Redact secrets and personal content, separate audit from debug logs, set retention and verify delivery during rotation/restart.

Required normalized fields: service, action, actor, record_count, result.

Illustrative collection event

Project-normalized synthetic JSON, not a native provider log, captured event, attack verdict or validated detection. A parser/adapter is required for a real SIEM. Example names, IPs and values are fictional.

{
  "schema": "1200km.telemetry.example.v1",
  "synthetic": true,
  "timestamp": "2026-09-27T12:00:00Z",
  "telemetry_id": "DC0038",
  "collector": "illustrative-lab-collector",
  "observation": {
    "service": "lab-portal",
    "action": "export_requested",
    "actor": "lab-reader",
    "record_count": 3,
    "result": "authorized"
  }
}

Visibility limits: Debug output is not necessarily a durable audit trail. Application developers must emit the relevant event; installing a log pipeline does not make an absent event exist.

File Access · source-mapped

DC0055 collection page — Enable Audit File System plus a narrow SACL on a lab directory; on Linux select path/syscall rules and join associated Audit records.

Providers and collection configuration

  • Use a lab-only GPO under Computer Configuration → Windows Settings → Security Settings → Advanced Audit Policy Configuration. Enable the specific subcategory named below.
  • For object-level auditing, configure a narrow system access control list (SACL) on the test object in addition to policy. Do not audit an entire production directory by default.
  • Forward Security events from the responsible host or controller; retain subject and target identities, object identifiers, result and original XML. Confirm effective policy with auditpol /get /category:*.

Required normalized fields: path, access, actor, result.

Illustrative collection event

Project-normalized synthetic JSON, not a native provider log, captured event, attack verdict or validated detection. A parser/adapter is required for a real SIEM. Example names, IPs and values are fictional.

{
  "schema": "1200km.telemetry.example.v1",
  "synthetic": true,
  "timestamp": "2026-09-27T12:00:00Z",
  "telemetry_id": "DC0055",
  "collector": "illustrative-lab-collector",
  "observation": {
    "path": "C:\\Lab\\demo.txt",
    "access": "read",
    "actor": "LAB\\analyst",
    "result": "success"
  }
}

Visibility limits: A policy checkbox alone does not guarantee an event. Host role, SACL, success/failure selection and audit policy precedence matter. Directory changes may arrive as multiple records.

Network Connection Creation · source-mapped

DC0082 collection page — Enable scoped Sysmon NetworkConnect (3), disabled by default, or a platform-equivalent process-aware network sensor.

Providers and collection configuration

  • Microsoft Sysmon: Windows event-based collection; enable the event types needed below.
  • Linux Audit: Linux alternative for supported system-call and file events; different semantics and fields.
  • Apple Endpoint Security clients: macOS alternative where the subscribed event exists; requires an entitled, approved client, not an iOS collector.
  • On a disposable Windows host, review the installed Sysmon schema and current configuration; merge scoped event filters into the existing policy rather than replacing it.
  • Forward Microsoft-Windows-Sysmon/Operational to the lab collector. Preserve event ID, timestamp, computer, process identifiers and the original event.
  • For Linux or macOS, select the equivalent sensor separately and test its emitted fields; a Windows event ID does not transfer across platforms.

Required normalized fields: process_guid, destination_ip, destination_port, protocol, initiated.

Illustrative collection event

Project-normalized synthetic JSON, not a native provider log, captured event, attack verdict or validated detection. A parser/adapter is required for a real SIEM. Example names, IPs and values are fictional.

{
  "schema": "1200km.telemetry.example.v1",
  "synthetic": true,
  "timestamp": "2026-09-27T12:00:00Z",
  "telemetry_id": "DC0082",
  "collector": "illustrative-lab-collector",
  "observation": {
    "process_guid": "lab-process-001",
    "destination_ip": "198.51.100.20",
    "destination_port": 443,
    "protocol": "tcp",
    "initiated": true
  }
}

Visibility limits: Event filtering and sensor versions change coverage. High-volume image-load and process-access collection needs tuning; endpoint events alone do not establish intent.

Process Creation · source-mapped

DC0032 collection page — Collect Sysmon ProcessCreate (1); preserve ProcessGuid and parent identifiers. Security 4688 is an alternative with separate audit/command-line settings.

Providers and collection configuration

  • Microsoft Sysmon: Windows event-based collection; enable the event types needed below.
  • Linux Audit: Linux alternative for supported system-call and file events; different semantics and fields.
  • Apple Endpoint Security clients: macOS alternative where the subscribed event exists; requires an entitled, approved client, not an iOS collector.
  • On a disposable Windows host, review the installed Sysmon schema and current configuration; merge scoped event filters into the existing policy rather than replacing it.
  • Forward Microsoft-Windows-Sysmon/Operational to the lab collector. Preserve event ID, timestamp, computer, process identifiers and the original event.
  • For Linux or macOS, select the equivalent sensor separately and test its emitted fields; a Windows event ID does not transfer across platforms.

Required normalized fields: process_guid, image, parent_image, command_line, user.

<EventFiltering>
  <ProcessCreate onmatch="include">
    <Image condition="is">C:\Windows\System32\whoami.exe</Image>
  </ProcessCreate>
</EventFiltering>

This fragment only selects the named process. It is not a complete production policy. Inspect the installed schema with sysmon64 -s; preserve existing rules, then apply the reviewed complete configuration with sysmon64 -c <configuration-file>.

Illustrative collection event

Project-normalized synthetic JSON, not a native provider log, captured event, attack verdict or validated detection. A parser/adapter is required for a real SIEM. Example names, IPs and values are fictional.

{
  "schema": "1200km.telemetry.example.v1",
  "synthetic": true,
  "timestamp": "2026-09-27T12:00:00Z",
  "telemetry_id": "DC0032",
  "collector": "illustrative-lab-collector",
  "observation": {
    "process_guid": "lab-process-001",
    "image": "C:\\Windows\\System32\\whoami.exe",
    "parent_image": "C:\\Windows\\System32\\cmd.exe",
    "command_line": "whoami",
    "user": "LAB\\analyst"
  }
}

Visibility limits: Event filtering and sensor versions change coverage. High-volume image-load and process-access collection needs tuning; endpoint events alone do not establish intent.

Script Execution · source-mapped

DC0029 collection page — Collect PowerShell 4104 with script-block ID and fragment sequence; instrument other engines separately rather than assigning them the same event ID.

Providers and collection configuration

  • Enable Script Block Logging through the policy/configuration for the installed edition. Forward Microsoft-Windows-PowerShell/Operational for Windows PowerShell, or PowerShellCore/Operational for PowerShell 7.
  • Retain script-block ID and fragment ordering alongside process creation. Protect access to collected script text; secrets may be included.
  • On Linux, use narrowly scoped Audit execution rules for the lab account and supported architectures, and retain joined SYSCALL/EXECVE/PATH records.

Required normalized fields: engine, script_block_id, fragment, fragments_total, script_text.

Get-WinEvent -FilterHashtable @{
  LogName = 'Microsoft-Windows-PowerShell/Operational'
  Id = 4104
} -MaxEvents 5

Read-only check after Script Block Logging is configured. For PowerShell 7 use its registered PowerShellCore/Operational channel. Missing events are not proof of no execution; check policy, channel and permissions.

Illustrative collection event

Project-normalized synthetic JSON, not a native provider log, captured event, attack verdict or validated detection. A parser/adapter is required for a real SIEM. Example names, IPs and values are fictional.

{
  "schema": "1200km.telemetry.example.v1",
  "synthetic": true,
  "timestamp": "2026-09-27T12:00:00Z",
  "telemetry_id": "DC0029",
  "collector": "illustrative-lab-collector",
  "observation": {
    "engine": "PowerShell",
    "script_block_id": "lab-script-1",
    "fragment": 1,
    "fragments_total": 1,
    "script_text": "Write-Output \"LAB_TELEMETRY_CHECK\""
  }
}

Visibility limits: Script content and process command lines are different signals. Logging may be partial, fragmented, filtered or disabled; never assume a command line reveals everything executed.

Detection rules and analytic guidance

Open the dedicated detection workspace · Existing query engineering

DET0927 Detect AI-Enabled Exfiltration

MITRE detection strategy

AN2079 Analytic 2079

An analyst observes an AI-associated process access data outside $AuthorizedAIDataScope$, followed within $EndpointExfilWindow$ by transfer-capable execution and an outbound connection to a destination outside $AuthorizedAIEndpoints$ from the same process lineage.

Platforms: Windows.

  • File Access — WinEventLog:Security; channel EventCode=4663, 4670, 4656.
  • Script Execution — WinEventLog:PowerShell; channel EventCode=4103, 4104, 4105, 4106.
  • Process Creation — WinEventLog:Sysmon; channel EventCode=1.

AN2078 Analytic 2078

An analyst observes an AI agent access $ProtectedDataScope$, invoke a tool/API using that information or a derived payload, and the destination service record a successful transfer to a recipient, repository, tenant, domain, or service outside $AuthorizedAIDataFlows$.

Platforms: SaaS.

AN2080 Analytic 2080

An analyst observes an AI-associated process access data outside $AuthorizedAIDataScope$, followed within $EndpointExfilWindow$ by transfer-capable execution and an outbound connection from the same lineage to a destination outside $AuthorizedAIEndpoints$.

Platforms: Linux.

Source rule directory and observation requirements

No exact-tagged Sigma source rule was imported. This does not establish that the technique is undetectable; the MITRE guidance and collection requirements remain relevant.

Before treating a rule as coverage

  1. Read the actual condition and logsource; identify attacker-side, victim-side or third-party visibility.
  2. Map native sensor fields, parse times and identities, then compile for the selected backend.
  3. Replay a behavior-positive case, a normal control, an authorized look-alike and a missing-telemetry case.
  4. Record matches and misses with rule version and source events. No match is inconclusive when required telemetry is absent.

Anomaly-based detection

No exact technique-specific Atlas model is present for T1700 in the selected source snapshot. The following is a design worksheet, not a newly validated model.

Anomaly engineering worksheet

Start from the exact analytic above. The collection-specific lenses below are analyst recommendations; they do not assert that an arbitrary statistical deviation detects this technique.

  • Application and service audit: Per application, client and route: request rate, rare actions, failure ratios and unusual state transitions. Version changes, retries, health probes and seasonal load need separate baselines.
  • Windows and directory audit: Per principal, host and object: unusual access/request rates, new principal-object relationships, and changes outside approved administration windows. Control for service accounts and scheduled directory maintenance.
  • Endpoint activity: Per host role and user: rare process/object combinations, bursts of object access or modification, and ordered parent/action sequences. Compare maintenance windows and signed software rollout activity.
  • Script and command execution: Per host role and principal: rare interpreters, command argument patterns and parent-child sequences. Normalize scripts carefully; obfuscation, administrative automation and missing command lines change visibility.
  1. Define an entity/peer group and feature; exclude the evaluation period from baseline training.
  2. Use representative normal and maintenance activity. Do not invent a production threshold from a synthetic example.
  3. Evaluate held-out behavior-positive and benign cases; measure false positives, misses and telemetry loss separately.
  4. Keep abnormality separate from maliciousness. Correlate with the behavior-specific source evidence.

Anomaly Detection Atlas · Baseline, validation and blind spots

Attack tools and documented software

No reviewed lab-tool association is selected for this exact technique. This is a curation gap, not a statement that no tool can implement the behavior.

Simulation design and documented procedures

Atomic evidence: no compatible candidate in this snapshot. Live lab status: not run for these imported procedures.

Use disposable lab hosts and test accounts, explicit target allowlists, a clean snapshot, bounded egress and a restore plan. Review every candidate and dependency before any execution.

No compatible Atomic procedure is present in this pinned evidence set. Technical feasibility has not been ruled out. A custom procedure must define the behavior, target, collection point, success observable and rollback before execution.

Technique-specific experiment contract

  1. Choose one procedure implementing T1700 AI-Enabled Exfiltration; identify the observable behavior in the source definition, not just its tactic.
  2. Record exact target, tool/version, inputs, privileges, network limits and approval. Never run an imported default target automatically.
  3. Enable the collection sources above and verify a benign baseline reaches the detector.
  4. Run only the approved isolated scenario; capture native events and an independent behavior observable.
  5. Evaluate the exact rule/model, restore the lab, and preserve configuration, event hashes, alerts, controls and misses.

Synthetic logs and offline contract check

Download synthetic collection fixture (JSON) · Download the complete engineering workbook

This fixture exercises the listed collection fields. It does not reproduce AI-Enabled Exfiltration, label its example events as malicious, or measure detection effectiveness. Use the exact analytic and selected procedure to build an independently labelled behavior-positive dataset.

Not run in this browser. No network scan or attack command is executed by this control.

Validation status and remaining work

Source identity
Joined by exact technique/object IDs against hash-pinned sources.
Collection fixture
Available for offline schema exercises; not attack emulation.
Native telemetry / backend / real attack
Not validated for these imported procedures. Source fidelity and page tests cannot establish sensor or detector effectiveness.

All-page audit and evidence inventory · This page’s machine-readable workbook

Sources and evidence boundaries

Pinned MITRE STIX source · SHA-256 81113ba98afa2928f3dcff62022539f7ce30c315c32e14422f11ff95ce534352. Definition and procedure context are reproduced from this snapshot, not independently re-investigated incidents.

Source-linked mitigations

Connected ecosystem references

Linked tags

Detection and collection

T1700 detection workspace

No reviewed association in this snapshot.

Existing research

Threat Matrix: knowledge routes, evidence and actor context

Pinned research references. No browser attack runner or production-validated detector is asserted. Imported procedures remain unvalidated; any bounded lab evidence has its own scope. ATT&CK / Atomic provenance · Detection provenance.