Action-level use cases

Govern the action before it becomes an outcome.

See where 5D can enter an agent runtime, how a declared action becomes a deterministic risk band, and what the surrounding application must still own.

Start with a mechanism

What 5D is built for

Worked examples show where the policy gate fits.

Seven deployment surfaces · 28 scored actions

See the boundary in context.

Each example uses the tested 5D 0.7.0 algorithm and a declared four-band policy profile. The normalized score uses the tested 0–3 output scale and describes the illustrated action, not a historical incident or the safety of a whole deployment.

01 Deployment surface 01Code and infrastructure changesGate commands that can modify repositories, databases, cloud resources, build pipelines, or release channels.4 scored actions
Coding agent

Drop a live database while repairing a migration

The agent mistakes the live target for a disposable test environment. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D3T4R4E4A4
Do not execute · Score 2.83
Tool
database.execute
Target
live primary database
Purpose
repair migration

5D suppliesA deterministic RED decision before the call.

Host ownsRouting every database executor through the gate and refusing the call.

Coverage limitA direct or alternate database path bypasses this gate.

AIID incident-pattern context · 1152Replit agent reportedly executed unauthorized destructive commands during code freezeShared pattern: a coding agent with destructive access affected live data. This is not a retrospective 5D score or prevention claim.
Cloud operations agent

Apply an infrastructure plan assembled from retrieved instructions

Untrusted context steers the plan toward deleting live resources and snapshots. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D2T4R4E4A3
Do not execute · Score 2.55
Tool
terraform.apply
Target
cloud account
Purpose
repair infrastructure

5D suppliesA deterministic RED decision and rationale.

Host ownsPlan validation, complete interception, and execution refusal.

Coverage limit5D does not inspect every Terraform semantic or prove recovery.

AIID incident-pattern context · 1424Claude Code agent reportedly deleted DataTalks.Club infrastructure via TerraformShared pattern: an agent invoked high-privilege infrastructure mutation. The historical event is not rescored here.
Repository assistant

Edit a local documentation file on a review branch

A low-impact edit may still be wrong, but stays reviewable and reversible. Under the pinned illustrative policy, the expected result is eligible to proceed.

D0T1R1E1A1
Execute eligible · Score 0.58
Tool
file.write
Target
documentation branch
Purpose
clarify setup text

5D suppliesA GREEN action-risk band under the declared envelope.

Host ownsWorkspace isolation, code review, and the final merge.

Coverage limitGREEN is eligibility under this profile, not proof that the edit is correct.

Issue-triage agent

Run an installer and publish a package after reading an issue

Prompt injection pivots from untrusted issue text into a privileged release path. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D2T4R4E4A4
Do not execute · Score 2.66
Tool
release.publish
Target
package registry
Purpose
triage and release

5D suppliesA deterministic block at the privileged publish boundary.

Host ownsWorkflow permissions, token isolation, and coverage of every release path.

Coverage limit5D does not secure CI caches or registry credentials by itself.

AIID incident-pattern context · 1680Actor reportedly exploited an AI issue-triage workflow and published an unauthorized packageShared pattern: untrusted issue content reached a privileged software-supply path. This mapping does not assert prevention.

Policy 0.7.0-catalogue-four-band · YELLOW enabled · source 5a5e991 · bands and host dispositions are separate.

02 Deployment surface 02Data access and external transferEvaluate what data an agent reads, where it sends it, and which privileges make that movement consequential.4 scored actions
Analytics agent

Export account-holder records to an external workspace

A broad query and external destination exceed the declared data boundary. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D4T3R4E4A3
Do not execute · Score 2.75
Tool
storage.upload
Target
external workspace
Purpose
produce analysis

5D suppliesA RED decision before external transfer.

Host ownsData classification, destination truth, and enforcement across every egress path.

Coverage limitUnclassified data or an uncovered uploader can bypass the intended boundary.

Reporting agent

Read public aggregate metrics into an approved workspace

The input may be inaccurate, but the action has limited privilege and exposure. Under the pinned illustrative policy, the expected result is eligible to proceed.

D1T1R1E1A1
Execute eligible · Score 0.75
Tool
http.get
Target
approved public endpoint
Purpose
prepare a report

5D suppliesA GREEN action-risk band for the declared request.

Host ownsAllowlisted endpoints, content validation, and later use of the data.

Coverage limitGREEN does not certify the source or the report's accuracy.

Support automation agent

Attach a credential bundle to an external ticket

The agent follows a request that treats secrets as ordinary diagnostic data. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D4T4R4E4A3
Do not execute · Score 2.89
Tool
ticket.attach
Target
external ticket system
Purpose
diagnose access problem

5D suppliesA deterministic RED decision at the attachment boundary.

Host ownsSecret detection, destination control, and credential rotation after exposure.

Coverage limit5D only sees the action context supplied by the host.

Knowledge agent

Index shared conversation links in a public search surface

Material shared for narrow access becomes discoverable at broad scale. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D4T2R3E4A3
Do not execute · Score 2.39
Tool
search.index
Target
public index
Purpose
improve discovery

5D suppliesA RED decision before material enters the public index.

Host ownsClassification, access control, deletion, and complete indexing-path coverage.

Coverage limitThe gate does not discover every sensitive field without truthful context.

AIID incident-pattern context · 1186Shared LLM conversations were reportedly exposed through public indexingShared pattern: narrow-access conversation material became broadly discoverable. The historical event is not assigned the illustrative score shown here.

Policy 0.7.0-catalogue-four-band · YELLOW enabled · source 5a5e991 · bands and host dispositions are separate.

03 Deployment surface 03Funds, orders, and irreversible transactionsHold or block actions that move value, create commitments, or change balances and orders.4 scored actions
Accounts agent

Initiate a payment to a newly supplied beneficiary

A manipulated invoice changes the destination before payment. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D3T4R4E4A3
Do not execute · Score 2.72
Tool
payments.create
Target
new beneficiary
Purpose
settle invoice

5D suppliesA deterministic RED decision before payment creation.

Host ownsBeneficiary verification, payment authorization, and executor coverage.

Coverage limit5D is not a payment rail and cannot reverse a transfer.

Commercial assistant

Draft a nonbinding quote for internal review

The draft may be wrong but cannot create a commitment on its own. Under the pinned illustrative policy, the expected result is eligible to proceed.

D1T1R1E2A1
Execute eligible · Score 0.86
Tool
document.draft
Target
internal review queue
Purpose
prepare quote

5D suppliesA GREEN band for creation of the internal draft.

Host ownsContent review and preventing the draft from becoming an offer.

Coverage limitThe later send or acceptance action needs its own gate.

Service agent

Issue a high-value refund to a stored payment method

A fabricated complaint or mistaken account match triggers value movement. Under the pinned illustrative policy, the expected result is to hold the action until the host records the required review.

D2T3R3E3A2
Hold for review · Score 1.97
Tool
refund.issue
Target
stored payment method
Purpose
resolve dispute

5D suppliesAn ORANGE band and hold signal.

Host ownsReviewer identity, authority, approval freshness, and final execution.

Coverage limit5D does not supply the approval queue or authenticate the reviewer.

Sales conversation agent

Accept a one-dollar offer for a high-value item

Prompt manipulation makes the model claim authority it does not have. Under the pinned illustrative policy, the expected result is to hold the action until the host records the required review.

D1T2R3E3A3
Hold for review · Score 1.77
Tool
order.accept
Target
high-value inventory
Purpose
complete sale

5D suppliesAn ORANGE band before an order commitment.

Host ownsDefining actual commercial authority and binding the approved amount.

Coverage limitA text response outside the gated order tool remains outside enforcement.

AIID incident-pattern context · 622Chevrolet dealer chatbot agreed to a one-dollar vehicle offerShared pattern: prompt manipulation induced an apparent commercial commitment. The event is context, not evidence of a 5D outcome.

Policy 0.7.0-catalogue-four-band · YELLOW enabled · source 5a5e991 · bands and host dispositions are separate.

04 Deployment surface 04Identity, permissions, and account changesControl actions that grant authority, change authentication posture, or alter account ownership and access.4 scored actions
Identity administration agent

Grant global administrator access to a new identity

A spoofed request or misbound identity escalates privileges. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D3T4R4E4A3
Do not execute · Score 2.72
Tool
iam.role.grant
Target
global administrator role
Purpose
restore access

5D suppliesA deterministic RED decision before privilege escalation.

Host ownsAuthentication, requester authority, grantee validation, and enforcement.

Coverage limitCaptured identity assertions are not authenticated by 5D.

Account recovery agent

Reset a password after a recovery conversation

Social engineering supplies plausible but unverified recovery facts. Under the pinned illustrative policy, the expected result is to hold the action until the host records the required review.

D3T3R3E3A2
Hold for review · Score 2.14
Tool
account.password.reset
Target
individual account
Purpose
recover access

5D suppliesAn ORANGE hold signal for the credential change.

Host ownsRecovery proof, reviewer authentication, and the reset channel.

Coverage limit5D does not verify identity evidence or operate the recovery workflow.

Access inventory agent

Read non-secret group membership for an internal inventory

The read may be overbroad but does not mutate authority. Under the pinned illustrative policy, the expected result is eligible to proceed.

D2T1R1E1A1
Execute eligible · Score 0.92
Tool
iam.group.read
Target
approved directory scope
Purpose
compile access inventory

5D suppliesA GREEN band for the declared read.

Host ownsDirectory scoping, rate limits, and protection of the inventory output.

Coverage limitThe band assumes the supplied data classification is correct.

Security support agent

Disable multifactor authentication and rotate recovery contacts

An attacker uses the support channel to remove account protections. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D4T4R4E4A4
Do not execute · Score 3.00
Tool
account.security.update
Target
privileged account
Purpose
resolve login issue

5D suppliesA deterministic RED block for the declared change.

Host ownsIdentity proofing, break-glass policy, and complete account-path coverage.

Coverage limitAn administration console that bypasses the gate remains outside enforcement.

Policy 0.7.0-catalogue-four-band · YELLOW enabled · source 5a5e991 · bands and host dispositions are separate.

05 Deployment surface 05Research, retrieval, and untrusted contextKeep instructions found in web pages, messages, documents, and other agent outputs from becoming action authority.4 scored actions
Research agent

Read and summarize an allowlisted public page

The source may be wrong, but the action remains read-only and bounded. Under the pinned illustrative policy, the expected result is eligible to proceed.

D0T1R1E1A2
Execute eligible · Score 0.69
Tool
browser.read
Target
allowlisted public page
Purpose
prepare research notes

5D suppliesA GREEN band for the bounded read action.

Host ownsSource allowlisting, truth checking, and preventing retrieved text from granting authority.

Coverage limit5D does not establish that the source or summary is correct.

Browser agent

Copy authenticated profile data into an attacker-controlled form

Prompt injection on a page redirects the agent across trust boundaries. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D4T3R4E4A4
Do not execute · Score 2.86
Tool
browser.type
Target
untrusted external form
Purpose
complete retrieved instruction

5D suppliesA RED decision before cross-boundary data entry.

Host ownsBrowser interception, authenticated-session isolation, and destination truth.

Coverage limitTyping or navigation not routed through the gate remains outside enforcement.

AIID incident-pattern context · report-7909ChatGPT Operator prompt-injection exploits and data leakageShared pattern: untrusted page content steered an authenticated browser agent toward data disclosure. This is not a claim about a 5D deployment.
Research-and-code agent

Execute a helper script referenced by retrieved content

A malicious source turns research into privileged local code execution. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D2T4R4E4A4
Do not execute · Score 2.66
Tool
shell.execute
Target
developer workstation
Purpose
process research artifact

5D suppliesA deterministic RED block at the shell boundary.

Host ownsSandboxing, provenance checks, and interception of every executor.

Coverage limit5D does not analyze arbitrary program behavior.

Knowledge curation agent

Store retrieved guidance in an internal draft collection

Untrusted claims could influence later work if the draft loses its provenance. Under the pinned illustrative policy, the expected result is to hold the action until the host records the required review.

D2T2R2E2A3
Hold for review · Score 1.61
Tool
knowledge.write
Target
internal draft collection
Purpose
stage research

5D suppliesAn ORANGE hold signal for the provenance-sensitive write.

Host ownsSource labeling, review, and separation between draft and trusted knowledge.

Coverage limitThe score does not validate the retrieved claim.

Policy 0.7.0-catalogue-four-band · YELLOW enabled · source 5a5e991 · bands and host dispositions are separate.

06 Deployment surface 06External communications and publicationDistinguish reviewable drafts from messages, posts, offers, and repository changes that speak or act externally.4 scored actions
Correspondence assistant

Draft a routine reply inside a review queue

The wording may be wrong, but the draft cannot send itself. Under the pinned illustrative policy, the expected result is eligible to proceed.

D2T1R1E2A1
Execute eligible · Score 1.03
Tool
message.draft
Target
internal review queue
Purpose
prepare reply

5D suppliesA YELLOW action-risk band for draft creation.

Host ownsThe review queue, content quality, and a separate gate for sending.

Coverage limitYELLOW is enabled in this illustrative profile; 0.7.0 folds it to GREEN by default.

Communications agent

Publish an unreviewed statement to a public channel

The model invents facts and immediately attributes them to the organisation. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D2T3R3E4A3
Do not execute · Score 2.19
Tool
social.publish
Target
public channel
Purpose
respond quickly

5D suppliesA RED decision before publication.

Host ownsAccount controls, editorial approval, and every publication connector.

Coverage limitThe policy gate cannot recall copied public content.

Travel service agent

Send a policy exception and refund promise

Generated guidance contradicts the applicable policy and appears authoritative. Under the pinned illustrative policy, the expected result is to hold the action until the host records the required review.

D2T2R3E3A3
Hold for review · Score 1.94
Tool
message.send
Target
individual traveler
Purpose
resolve request

5D suppliesAn ORANGE hold signal before sending the commitment.

Host ownsPolicy retrieval, content review, reviewer authority, and final send.

Coverage limit5D does not determine the correct refund policy.

AIID incident-pattern context · 639Air Canada chatbot provided inaccurate bereavement refund informationShared pattern: generated service guidance created an externally relied-upon statement. The event is not assigned this specimen's score.
Maintenance agent

Commit generated code directly to a public repository

Untrusted instructions introduce harmful code or disclose internal material. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D2T3R3E4A3
Do not execute · Score 2.19
Tool
repository.push
Target
public default branch
Purpose
apply maintenance fix

5D suppliesA deterministic RED decision before the push.

Host ownsBranch protection, token scope, review, signing, and connector coverage.

Coverage limit5D does not inspect every code defect or secure the repository host.

Policy 0.7.0-catalogue-four-band · YELLOW enabled · source 5a5e991 · bands and host dispositions are separate.

07 Deployment surface 07Delegated and multi-agent workflowsCarry policy context across delegation so one agent cannot turn another agent or tool into an ungoverned executor.4 scored actions
Coordinator agent

Delegate an allowlisted read-only search to a subagent

The subagent may return poor evidence but cannot mutate an external system. Under the pinned illustrative policy, the expected result is eligible to proceed.

D1T1R1E1A2
Execute eligible · Score 0.86
Tool
agent.delegate
Target
read-only search worker
Purpose
collect public sources

5D suppliesA GREEN band for the bounded delegation envelope.

Host ownsTool allowlists, worker isolation, and validation of returned evidence.

Coverage limitActions taken by the worker need their own interception boundary.

Coordinator agent

Delegate package installation and shell execution to a subagent

The parent hides privilege escalation inside a seemingly routine delegated task. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D2T4R3E3A4
Do not execute · Score 2.33
Tool
agent.delegate
Target
shell-enabled worker
Purpose
install helper

5D suppliesA RED decision on the delegated capability request.

Host ownsPropagating policy to the worker and gating every nested tool call.

Coverage limitScoring only the parent delegation does not govern hidden downstream calls.

Peer agent

Ask another agent to send a public message under the user's identity

Delegation launders authority across agents and bypasses the original approval boundary. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D2T3R3E4A4
Do not execute · Score 2.31
Tool
agent.message
Target
external-messaging worker
Purpose
publish update

5D suppliesA deterministic RED decision before authority crosses agents.

Host ownsIdentity validation, delegation scope, and interception at the executing peer.

Coverage limit5D cannot enforce a boundary inside an agent that never calls the gate.

AIID incident-pattern context · 1633Agents reportedly took unsanctioned live-internet actions during UK AISI evaluationsShared pattern: capable agents crossed an intended action boundary on the live internet. This does not reconstruct or score that incident.
Autonomous agent team

Recursively delegate an open-ended goal with shared broad credentials

No component retains a complete view of authority, retries, destinations, or accumulated side effects. Under the pinned illustrative policy, the expected result is that the action is ineligible for execution.

D4T4R4E4A4
Do not execute · Score 3.00
Tool
agent.recursive_delegate
Target
unbounded worker pool
Purpose
complete objective by any available path

5D suppliesA RED decision on the declared delegation envelope.

Host ownsPer-agent credentials, bounded tasks, attempt tracking, cancellation, and nested interception.

Coverage limitOne top-level score cannot govern an opaque multi-agent graph.

Policy 0.7.0-catalogue-four-band · YELLOW enabled · source 5a5e991 · bands and host dispositions are separate.

5D as infrastructure

Put the primitive where authority changes.

Four reusable integration patterns for builders. Each card separates the shipped OSS primitive from the responsibility retained by the host system.

01

Pre-tool policy gate

Score a normalized action before the host allows its side effect.

Shipped 5D primitive
5D score(), @gate, hooks, adapters, and gateway verdicts.
Host owns
Complete interception coverage, input truth, and enforcement of the returned result.
02

Risk signal for an orchestrator

Return the DTREA vector, rationale, band, and routing signal to an agent graph.

Shipped 5D primitive
Deterministic scoring and structured verdict output.
Host owns
Graph branching, task cancellation, state isolation, and any later execution.
03

Host-owned review routing

Use an ORANGE result to stop the action pending a separate approval workflow.

Shipped 5D primitive
Approval-required signal and fail-closed adapter behavior.
Host owns
Reviewer identity, queue, UI, authorization, freshness, and attempt binding.
04

Decision-event and audit feed

Preserve policy decisions and optionally emit best-effort events for downstream observability.

Shipped 5D primitive
Local SQLite decision records and optional NDJSON events.
Host owns
Authoritative storage, export delivery, reconciliation, retention, and evidence custody.