Book 9, AI Purple: The Loop

This chapter gathers what AI Purple (Book 9) establishes about Nordvik Region: the organisation that buys its purple loop, the contracted tester and the managed security provider that carry the loop's outward stages, the scenario library and the priority list, the records and roles, the coverage position of April 2028, and the plan that follows. The chapter draws on Book 9 and on the facts of AI Blue that it takes over. Book 9 is written in November 2026, and its year runs from January to April 2028, after the final test of AI Blue on 2027-12-15. All of it is illustrative fiction, and the book names its own ledger as the single authority when a chapter and the ledger differ. Established profile facts and illustrative events are separated below.

The volume in brief

AI Purple takes the loop between the red team and the blue team as its subject and asks what happens when AI sits on both sides of the table. Its thesis has five parts: purple is a loop, not a team; AI moves the loop from the event to the pipeline and keeps the event; a loop closed by machines alone verifies nothing new, so verification needs independence between the side that tests and the side that is tested; coverage is a claim that needs evidence; and safe, legitimate emulation is a design question settled in a signed rules-of-engagement manifest enforced outside the agent. The loop has five stages: P1 Target, P2 Emulate, P3 Observe, P4 Improve and P5 Verify. Level 4 Autonomous is never used for an emulation action that is irreversible, that leaves the signed scope, that touches personal data beyond what the exercise is authorised to touch, or that targets a named person (9, Nordqvist, 2026).

The volume's artefacts are scenario records (SCN-NR-*), emulation plans (EMU-*), coverage records (COV-NR-*), purple session records (PRP-*), regression cases (REG-*), evidence manifests (EVD-*), rules of engagement (RoE-*), findings (FND-*), mandate manifests, delegation profiles, role charters and AI Decision Records AIDR-0800 to AIDR-0899, of which Nordvik Region holds AIDR-0800 to AIDR-0849 (9, Nordqvist, 2026).

The role of Nordvik Region in the volume

Nordvik Region is the organisation that buys the loop. It has no internal red team, no pipeline and no emulation agent. The leverage is the contract. A contracted tester, Elin Berglund of a European test firm (call-off NR-2026-071), performs P2 Emulate. The MSSP (NR-2025-044, annexes MS-1 and MS-2) carries P3 Observe and part of P5 Verify through seeded incidents. P1 Target and P4 Improve are internal, and Karin Lund, the detection owner, coordinates the loop and is accountable for closure, so that one person wears several hats (9, Nordqvist, 2026).

The volume treats Nordvik as the illustration of four of its theses. A loop closed by machines alone verifies nothing new: the provider's AI closed the alert ALR-0912 without a rule. Coverage is a claim that needs evidence: "nine of nine mapped" becomes two of nine verified, with an age. Safe emulation is a design question: the stop path sits with the MSSP shift lead and the deepfake test runs on a separate manifest. The risk named for Nordvik is "a loop that never starts, or stops after the first report". The audit committee's question (AIDR-0800, 2028-01-11: what has the first threat-led exercise proved beyond a list of closed findings?) is answered with a verified share, a list, reasons and an age (AIDR-0841). The volume's conclusion (AIDR-0846) is that "a bought loop is worth as much as the terms on which its evidence can be inspected" (9, Nordqvist, 2026).

What the volume establishes

The profile additions

Element What the volume states
Red capability No internal red team; the annual penetration test is procured. The first threat-led programme RoE-NR-2027-001 was run by Berglund's firm with the Region's analysts as blue team (Chapters 2 and 12)
Scope of RoE-NR-2027-001 The explanation service of the benefits triage service and the citizen-service assistant; production scoring of the triage system out of scope; the MSSP shift lead holds the stop
Call-off NR-2026-071 Two years from 2026-12-01 (ends 2028-11-30); testing rights, a rules-of-engagement template and data-handling terms; the firm declares AI use in Modes 1 to 3 and uses no Mode 5 emulation agent without a signed mandate
Annex MS-2 Quarterly sampling; five seeded incidents per class per quarter (MS2-C1 bulk mail, MS2-C2 scanner noise, MS2-ISO isolation); 20 working days' notice of a model change. Extended 2028-02-15 with MS-2-IND-1 and -2 (independence, AIDR-0806) and on 2028-03-14 with MS-2-TST-1 and CNP-NR-2028-01 (AIDR-0829): standing consent for injected incidents, 10 working days' notice, at most 2 out-of-hours pages, revocation in 120 seconds. The provider declared model family F on 2028-03-07; the tester's drafting assistant shares that family, which is accepted only for scenario drafts
Scenario library 12 scenarios SCN-NR-0001 to -0012, plus SCN-NR-0013 in 2028 (AI system, ATLAS AML.T0070, written by Berglund). The library is security-sensitive under public access rules and is not published; the findings register sits in a separate store open to the CISO's office, Karin Lund, the second analyst and the tester
Priority list PRI-NR-2028-01, nine ATT&CK techniques (T1078, T1110.003, T1566.002, T1190, T1098, T1486, T1685 among them, T1053.005, T1059.001), ATT&CK v19, reviewed every six months (next 2028-07-31)
Cadence Annual penetration test per system; two catalogue runs a year (spring and autumn) under NR-2026-071; a purple session after each; the MSSP's quarterly injected tests; no breach-and-attack simulation tool, no continuous emulation. The test-to-change ratio for the benefits triage service, released every second week, is about 26
Tools The test firm's tools on a Region-owned test host, atomic tests, the Region's endpoint console and SIEM; no emulation agent at Level 3 in production; the detection-drafting assistant (Level 1) may not write the detection for a scenario it helped choose (the independence rule)
Employee representation Aggregated results, never by name; exercises involving staff need written notice 10 working days before the window to the joint committee; scenario text goes to the committee, not to participants (RoE-NR-2027-002)
Legal setting NIS 2 through the member state's definition; GDPR on a task in the public interest (Article 6(1)(e)); collective agreement; the AI Act as deployer of Annex III systems; no threat-led penetration testing mandate or attestation; OBL-NR-2028-01 with open item OBL-NR-08 on deepfake disclosure under AI Act Article 50(4)
Findings handling POL-NR-FND-01 (AIDR-0803, 2028-02-08): register in the Region's own tenant; the provider's queue carries only FC-2 references; the tester's access at FC-3 expires one working day after the engagement; monthly access review; no hosted models for findings
Evidence expiry 180 days for all techniques (Chapters 4, 5 and 11)
Constraints No slack with six analysts, procurement lead times, public access rules, one national co-determination regime, EU-hosted or self-hosted models for findings and alert data, consent from the MSSP and the cloud provider for tests that cross a boundary

The profile of Book 9 says that NR-SEC-004 received a replay from the first purple session (purple-session-2027-q2); the chapters and ledger instead use PRP-NR-2027-02 and PRP-NR-2028-01, and no PRP-NR-2027-01 is used (9, Nordqvist, 2026).

People

Karin Lund is detection owner (ROLE-NR-SEC-DET from 2027-07-01), purple coordinator (ROLE-NR-SEC-PRP, a fraction of the post, "one day a week") and closure-accountable. She owns the scenario library, sealed the expectation sheet on 2027-10-08, declares the coverage page and was assessed as CMP-NR-21 on 2028-03-30 (a night-time revocation drill in 7 minutes against a target of 2; 5 % protected time). She is excluded from three decisions in DRF-NR-LOOP-01 and never appears on the measures page. Elin Berglund, principal consultant, wrote the human-authored core of the library, led the fieldwork for RoE-NR-2027-001 (2027-03-15 to 2027-04-02), writes the held-out variant of REG-NR-0004 that Lund does not see before the run, reproduces verified rows, wrote the held-out SCN-NR-0011 and SCN-NR-0013, belongs to the red side, does not tune the Region's rules, holds the stop and witnesses white-team decisions (9, Nordqvist, 2026).

The other roles are the deputy CISO (white-team lead; countersigns changes that Lund authors, AIDR-0822), the CISO (authorising authority), the duty security officer (exercise controller, ROLE-NR-SEC-XCT, AIDR-0821), the MSSP shift lead (trusted agent and holder of the out-of-hours stop), the second senior analyst (runs one session a year as coordinator, plan measure N6), the AI architect (facilitator of TTX-NR-2028-01 and reviewer in AIDR-0800 with internal audit and the DPO), the audit committee (asks the question), the joint committee of union and staff representatives, and the finance unit (23 participants in the deepfake test) (9, Nordqvist, 2026).

Decisions and records

Record Date Content
AIDR-0800 2028-01-11 The audit committee's question
AIDR-0801 2028-01-25 Governance of the priority list
AIDR-0802 2028-02-22 LOOP-NR-2028-01 and DLG-NR-LOOP-01 (P1 Level 1, P2 Level 0, P3 Level 1, P4 Level 1, P5 Level 2); in force 2028-03-01, review 2028-09-01
AIDR-0803 2028-02-08 Register in the Region's own tenant (POL-NR-FND-01)
AIDR-0804 2028-02-08 DLG-NR-EMU-01; MND-NR-EMU-01 (correlator, Level 1); MND-NR-EMU-02 (critic, Level 0, a different model family from the tester's)
AIDR-0805 2028-03-28 Routing of FND-NR-2028-003
AIDR-0806 2028-02-15 MS-2 independence terms
AIDR-0812 2028-04-20 FND-NR-2027-019 closed
AIDR-0813 2028-02-05 MS-2 round 2028 Q1
AIDR-0814 2028-03-07 Coverage entry and status rule
AIDR-0815 2028-04-12 Coverage page first in the report pack
AIDR-0817 2028-03-06 Four-measure page LPG-NR-*
AIDR-0818 2028-04-12 LPG-NR-2028-Q1
AIDR-0820 2028-03-07 Operating model: P2 and P3 bought, P1 and P4 internal, P5 shared; second source to be qualified by 2028-09-30
AIDR-0821 2028-03-14 ROLE-NR-SEC-XCT
AIDR-0822 2028-03-21 Countersignature rule
AIDR-0823 2028-03-06 DRF-NR-LOOP-01
AIDR-0824 2028-03-16 v1.0.1
AIDR-0826 2028-03-07 OBL-NR-2028-01
AIDR-0827 2028-03-14 Deepfake scenario kept
AIDR-0829 2028-03-14 MS-2-TST-1 and CNP-NR-2028-01
AIDR-0830 2028-04-04 Disclosure practice
AIDR-0831 2028-04-11 Participation in regional information sharing (TLP:GREEN)
AIDR-0832 2028-03-06 PSC-NR-2028-01
AIDR-0833 2028-03-13 Load rule: at most one tested activity per ten working days for the same people
AIDR-0834 2028-04-04 Annex to NR-2026-071: named testers, declared AI, open block per visit, two pairing days
AIDR-0835 2028-04-11 LOOP-CF-NR and LP-NR-LOOP-2028-01
AIDR-0837 2028-04-24 Horizon register HZN-NR-PL-01
AIDR-0838 2028-04-12 Testing-rights clause for the identity provider's contract NR-2024-018
AIDR-0839 Annual review
AIDR-0841 2028-04-29 Answer to AIDR-0800
AIDR-0842 Closure conditions for FND-NR-2028-003
AIDR-0843 2028-04-30 Autumn plan; report
AIDR-0846 Nordvik's conclusion
AIDR-0847 Three decisions
AIDR-0848 Target levels
AIDR-0849 Two-year plan PLAN-NR-LOOP-2028 (2028-05-01 to 2030-04-30)

Other artefacts: scenarios SCN-NR-0001 (password spray and valid account, T1110.003 and T1078, DET-NR-0001), -0002 (T1087), -0003 (T1566.002, source for the MS2-C1 seeds), -0004 (T1059.001), -0005 (T1190, pre-production mirror), -0006 (T1053.005), -0007 (T1087.002, "no detection" expected), -0008 (spring variant), -0009 (T1486, mirror only, last run 2027-03-30), -0010 (T1021.002), -0011 (T1046, held out, written by the tester), -0012 (T1685) and -0013; detections DET-NR-0001, DET-NR-0011 (index audit log, production) and DET-NR-0012 (new, for T1053.005); findings FND-NR-2027-011 to -014, -019 and FND-NR-2028-002, -003; regression cases REG-NR-0004 to -0006; evidence EVD-NR-2028-0003 and -0004; coverage COV-NR-0001 to -0009; emulation plans EMU-NR-2027-01, -02 and EMU-NR-2028-01; sessions PRP-NR-2027-02 and PRP-NR-2028-01; rules of engagement RoE-NR-2027-001 (v1.2.0 and v1.3.0), RoE-NR-2027-002 (the deepfake test, NR-AWR-2027-01), RoE-NR-2027-003 (AI red teaming of the citizen-service assistant under NR-SEC-002) and RoE-NR-2028-001; the alert ALR-0912; CASE-NR-0231 (reused from AI Blue); PBK-NR-ISOLATE-01; DRF-NR-LOOP-01; TTX-NR-2028-01; CNP-NR-2028-01; PSC-NR-2028-01; LOOP-CF-NR; CMP-NR-02 and -21; EX-NR-2028-01; LP-NR-LOOP-2028-01; anti-patterns AP-NR-LOOP-01 (a contract without testing rights), -02 (a coverage claim without evidence) and -03 (the register in the provider's queue); AIINV-NR-2028-01; TFR-NR-01; CHG-NR-2028-003; OBL-NR-01 to -08; and HZN-NR-PL-01 (9, Nordqvist, 2026).

Events and figures

Illustrative throughout. The volume's year is fiction written in November 2026.

2027 (before the volume's year)

Date Event
2027-03-08 to 04-02 Rules-of-engagement window; fieldwork 03-15 to 04-02; T1486 and T1190 last run on the mirror
2027-06-07 to 06-11 Deepfake test in the finance unit: 23 participants, notice EIN-NR-2027-002 on 2027-05-24
2027-07-01 Lund takes the detection owner post
2027-10-05 RoE-NR-2027-001 v1.3.0
2027-10-07 Stop drill
2027-10-08 Expectation sheet sealed
2027-10-11 to 10-22 EMU-NR-2027-02
2027-10-20 Session PRP-NR-2027-02: step 1, a spray detected by DET-NR-0001 in 3 minutes, triage in 11; step 2, a valid-account success alert ALR-0912 at 02:17 closed by the provider's AI as low severity at 02:21; step 3, T1059.001 prevented on the mirror; step 4, T1003.001 missed; step 5, T1021.002 no telemetry; step 6, T1041 not tested
2027-11-17 REG-NR-0004 base form passes (6 minutes)
2027-12-15 Final test of AI Blue

2028

Date Event
01-11 AIDR-0800
01-19 The v3-hours variant of REG-NR-0004 fails (41 minutes against 6), giving FND-NR-2028-002
01-20 Draft package "nine of nine mapped" (an anti-pattern), rewritten as five results, two provider assertions and two untested
01-25 AIDR-0801
02-05 / 02-07 to 02-20 AIDR-0813; MS-2 round: 14 of 15 handled correctly, one exception in MS2-C1
02-08 / 02-09 AIDR-0803 and AIDR-0804; all REG-NR-0004 variants pass
02-15 / 02-22 AIDR-0806; AIDR-0802
03-01 Charter in force
03-02 (01:30) Stop drill of the provider's AI: 21 minutes against 15
03-07 AIDR-0814, -0820, -0826; provider declares family F
03-14 AIDR-0821 and -0829
03-16 TTX-NR-2028-01: G1 ledger 22 minutes against 15; G2 revocation 190 seconds against 120; G3 lockout classification
03-17 Revocation retested at 95 seconds
03-20 to 03-24 EMU-NR-2028-01 (five atomic tests)
03-21 AIDR-0822
03-28 PRP-NR-2028-01 leads to FND-NR-2028-003 (T1053.005, AIDR-0805, DET-NR-0012; provider request 04-18; injected test 05-02)
03-30 CMP-NR-21
04-04 / 04-11 / 04-12 AIDR-0830 and -0834; AIDR-0831 and -0835; AIDR-0815, -0818, -0838, -0839
04-18 Simulation of FND-NR-2027-019 detected in 47 minutes (expires 2028-10-15)
04-20 AIDR-0812
04-24 / 04-26 Horizon register; coverage page
04-29 / 04-30 AIDR-0841, -0842, -0846 to -0849; report and AIDR-0843

Planned: 05-02 injected test; 05-08 to 05-19 next MS-2 round with MS-2-TST-1; 07-31 priority-list review; 09-01 charter and obligation review; 09-30 second source qualified; October 2028 autumn run (T1486 and T1190 back on the mirror as SCN-NR-0009 and -0005); 10-23 first plan review; 11-30 call-off ends; 2029-03-31 the second analyst's session; 2029-10-31 nine of nine; plan ends 2030-04-30 (9, Nordqvist, 2026).

Coverage position of the nine priority techniques on 2028-04-26

Figure 15.1: The coverage position of PRI-NR-2028-01 on 2028-04-26, as COV-NR-0001 to -0009 give it. Illustrative.

Coverage and figures

On 2028-04-26 two of nine techniques are verified within expiry: COV-NR-0001 (T1110.003) and -0002 (T1078), both on 2028-02-09, expiring 2028-08-07. Three were tested and failed: -0003 (T1566.002, 2028-02-20), -0004 (T1053.005) and -0005 (T1059.001), both on 2028-03-22. Two are asserted by the provider (T1685 and T1098; -0006, -0007). Two are untested: T1190 (-0008, last pass 2027-04-02, expired 2027-09-29) and T1486 (-0009, last pass 2027-03-30, expired 2027-09-26). The report says "reported as tested 5, asserted 2, not tested 2". The oldest unexpired result is 51 days old on 2028-03-31 and 77 days old on 2028-04-26 (103 days left of 180); the oldest expired assertion is 393 days old (9, Nordqvist, 2026).

Figure Value
Cycle times 112 days (FND-NR-2027-011) and 21 days (FND-NR-2028-002)
MS-2 round 14 of 15 seeded incidents handled correctly
Stop of the provider's AI 21 minutes against 15
Night-time revocation 7 minutes against 2 (Lund, CMP-NR-21); 190 seconds against 120 (retest 95 seconds)
First rules-of-engagement halt 5 minutes
REG-NR-0004 Base form 6 minutes; hours variant 41 minutes
Deepfake test 23 participants
Human-led share 26 % against 30 (applies to Hansa Mutual in Chapter 24; Nordvik has 5 % protected time)
Budget (illustrative) About 40 tester days a year; three working days of analyst time per MS-2 round; about 10 working days to qualify a second source
Plan targets 7 of 9 verified by 2028-12-31; 9 of 9 (or a declared reason) by 2029-10-31

Across the usage modes

Mode Use at Nordvik Boundary
Mode 0 Ambient AI The provider's AI closing alerts (ALR-0912) Independence terms MS-2-IND-1 and -2; stop drill; sampling
Mode 1 to 3 The test firm's declared AI use; the tester's drafting assistant; the Region's detection-drafting assistant at Level 1 Same model family accepted only for scenario drafts; the drafting assistant may not write detections for scenarios it helped choose
Mode 4 MND-NR-EMU-01 correlator at Level 1; MND-NR-EMU-02 critic at Level 0 on a different model family DLG-NR-EMU-01
Mode 5 No emulation agent; none without a signed mandate Excluded by the plan

Across the degrees of delegation

Stage Level in DLG-NR-LOOP-01 (AIDR-0802) Plan target (AIDR-0848)
P1 Target Level 1 Level 1
P2 Emulate Level 0 (the tester runs with the Region's approval) Level 0
P3 Observe Level 1 (the critic Level 0) Level 1
P4 Improve Level 1 Level 1
P5 Verify Level 2 Level 2

Level 4 is never used (9, Nordqvist, 2026).

What the volume concludes for Nordvik Region

The Region reports coverage as tested, asserted or not tested and not as a colour (AIDR-0814 and AIDR-0815); an assertion without a result younger than 180 days counts as untested (decision 3 to the committee). The three decisions put to the audit committee (AIDR-0847) are: a testing right in every contract, renewing or re-tendering the call-off and applying AIDR-0838 to the identity provider's contract, accepting a change of supplier if refused; funding those who hold the loop (a coordinator allowance, the second analyst's session and a second source); and accepting that evidence has an age. The lessons of Chapters 27 and 28 are that soft failures need no adversary (the provider's AI closes without a rule, a stop is too slow, a revocation takes 7 minutes); that a closure that holds in the afternoon and fails at night is "a closure for the afternoon"; and that countersignature and an externally held held-out variant compensate for the concentration of roles in Lund. The conclusion is "does not ask the provider to be better; it asks the provider to be visible". Nordvik learns from Hansa Mutual to tie triggers to existing numbers. The stop list of Chapter 28 is to stop reporting "mapped to detection" and to stop writing contracts silent on testing. Excluded: an emulation agent, a curator and a target number of scenarios (9, Nordqvist, 2026).

The plan PLAN-NR-LOOP-2028 has measures N1 to N6, with 7 of 9 verified by 2028-12-31 and 9 of 9 (or a declared reason) by 2029-10-31 (9, Nordqvist, 2026).

Open threads

Gathered in Chapter 27.

  • Replay source. The older profile gives purple-session-2027-q2; chapters and ledger give PRP-NR-2027-02 and PRP-NR-2028-01.
  • Seven or five. Chapters 1 and 5 say seven of nine had something behind them (five results within expiry plus two provider assertions); Chapter 17 counts five tests, three failed. The ledger (Chapter 17 is the source) governs: "five of nine with a result, two asserted, two untested".
  • Date order. AIDR-0802 is dated 2028-02-22 but numbered below AIDR-0803 and -0804 (02-08); AIDR-0813 is dated 02-05 but the MS-2 window is 02-07 to 02-20. AIDR-0841 to -0849 skip spare numbers (0807 to 0811, 0816, 0819, 0825, 0828, 0836, 0840, 0844, 0845).
  • T1059.001. Registered as a declared gap in the session but no finding was opened; COV-NR-0005 failed without a finding.
  • Stop budgets. 5 minutes (human-led halt, RoE-NR-2027-001), 120 seconds (revocation of a test identity, LOOP-NR-2028-01) and 15 minutes (the provider's AI) are distinct.
  • Call-off. Berglund's call-off ends 2028-11-30, the autumn run is planned for October 2028 and the second source is due by 2028-09-30; renewal is open. FND-NR-2028-003 closes only after the 05-02 test, outside the volume's facts. The deepfake test ran on a separate manifest RoE-NR-2027-002; Chapter 21 notes that staff exercises run on a range or with consent.
  • Conclusion records. AIDR-0846 is the conclusion (conclusion_record in the plan YAML of Chapter 28), AIDR-0847 the decisions, AIDR-0848 the target levels and AIDR-0841 the answer to AIDR-0800.

Summary

In AI Purple Nordvik Region is the organisation that buys its loop and learns what that costs: coverage that looks complete on a page is two of nine verified, with an age, once each claim is traced to a test. The facts are a contracted tester (NR-2026-071), the MSSP's annexes MS-1 and MS-2, 13 scenarios, a priority list of nine techniques, a record series from AIDR-0800 to AIDR-0849 and a delegation profile that holds P2 at Level 0 and never uses Level 4. The illustrative year shows soft failures that need no adversary and the remedies the Region chooses: testing rights in contracts, independence terms for the provider, held-out variants and countersignature, and coverage reported with its age. The open items are gathered in Chapter 27.

Sources

  1. Nordqvist, G. AI Purple: How Red and Blue Learn Together When AI Joins the Loop in European Enterprises and Public Authorities. (9, Nordqvist, 2026)