deep-hunt

name	deep-hunt
description	Manual workflow attack mode for authorized bug bounty and pentest work. Use as /deep-hunt <workflow> to attack one business workflow deeply instead of scanning endpoints shallowly.
sources	field_recon, hackerone_public
report_count	50

Deep Hunt Mode

Use this mode when the operator invokes /deep-hunt <workflow>.

Supported workflows:

• recovery
• invite
• billing
• graphql
• import
• export
• roles
• support
• ai
• race
• oauth
• saml
• upload
• webhook
• generic

If the workflow is not recognized, map it to the closest one and state the mapping.

Output Contract

Produce:

• workflow
• crown_jewel_alignment
• actors_needed
• object_families
• state_machine
• attack_matrix
• side_effects_to_chase
• first_10_manual_tests
• chain_candidates
• evidence_to_capture
• stop_conditions

The output must be actionable enough that the operator can run it in Burp/Caido immediately.

Universal Deep-Hunt Loop

For the chosen workflow, test six boundaries:

1. Actor boundary: owner, member, invited, removed, expired, unauthenticated, API token, mobile session, support-like role.
2. Object boundary: primary ID, child ID, attachment ID, export ID, job ID, GraphQL GID, signed URL, webhook ID, AI source ID.
3. State boundary: draft, pending, approved, rejected, archived, deleted, imported, shared, transferred, expired, refunded, downgraded.
4. Shape boundary: JSON/form/multipart, scalar/array/object, duplicate keys, null plus real value, method override, mobile headers, content negotiation.
5. Time boundary: stale token, queued worker, after access removal, after role downgrade, during retry, parallel sessions.
6. Side-effect boundary: export, notification, webhook, audit log, support view, AI citation, background job output, generated file.

Do not declare the workflow done until all six have at least one concrete test result.

Workflow Playbooks

recovery / invite

Prioritize:

• destination field confusion
• stale link redemption
• target identity switch between generation and redemption
• preview/accept/resend/cancel mismatch
• recovery side channels in notifications, logs, webhooks, support views

First tests:

• scalar email -> array/object
• duplicate destination keys
• generate token for attacker, switch target state, redeem stale token
• accept invite after revocation
• compare web/mobile/API redemption paths

billing

Prioritize:

• archived prices
• coupon/credit/rate races
• refund without entitlement revocation
• webhook replay/type confusion
• seat count and plan downgrade edges

First tests:

• finalize after cancellation
• apply discount after invoice finalization
• refund while consuming entitlement
• parallel redemption of same value
• stale idempotency key reuse

graphql

Prioritize:

• object-family authorization mismatch
• helper mutation drift
• mixed-tenant nested IDs
• alias/batch partial leaks
• read/mutate mismatch

First tests:

• enumerate IDs/GIDs by family
• pair attacker parent with victim child
• pair victim parent with attacker child
• run read, preview, export, mutate, delete for same object family
• batch allowed and forbidden objects together

import / upload

Prioritize:

• parser and archive trust
• symlink/hardlink/duplicate-name handling
• import from URL / redirect / local path confusion
• metadata and filename sinks
• imported object ACL drift

First tests:

• symlink archive into upload/import lane
• duplicate filenames and nested archives
• source URL redirect to internal-like destination
• payloads in metadata, CSV formulas, markdown, diagrams
• import private object reference into attacker workspace

export

Prioritize:

• stale authorization in workers
• child-object direct access
• signed URL drift
• export requested by one actor and downloaded by another
• import/export ACL laundering

First tests:

• request export, remove access, download when job completes
• direct child ID download
• signed URL reuse after role downgrade
• export victim object via attacker-controlled parent/reference
• compare CSV/PDF/ZIP variants

roles / support

Prioritize:

• read-only role write side effects
• helper object ownership
• support/operator rendering sinks
• API key, webhook, queue, alert, saved-search creation
• support-visible lower-privileged data

First tests:

• create helper object as observer/support role
• configure export/webhook/notification without read permission
• grant/assign through helper object
• inject marker into field rendered in support/admin context
• compare UI-hidden actions with API-available actions

ai

Prioritize:

• cross-tenant source/citation leakage
• vector store and uploaded file IDOR
• saved prompt/tool config exposure
• action/tool misbinding
• support transcript leakage

First tests:

• swap thread/file/vector store/source IDs across tenants
• ask for citations and debug summaries
• duplicate assistant after access changes
• move/import/re-index documents and retest retrieval
• inspect support/admin views of AI conversations

race

Prioritize:

• check-then-act gaps
• duplicate value issuance
• stale worker authorization
• retry/replay and idempotency mistakes
• parallel state transitions

First tests:

• parallel requests from two sessions
• approve/delete, invite/revoke, pay/cancel, refund/consume
• replay webhook and idempotency keys
• send browser and API request simultaneously
• observe delayed workers and retry queues

Completion Standard

End with:

• the strongest current primitive
• the best next pivot
• what proof is missing for reportability
• whether /attack-chain should be invoked now