Security, Operations and Resilience

This chapter gathers the security, operations and resilience posture of Hansa Mutual as the volumes build it up: who runs security and operations, what the group owns and what it buys in blue, red and purple capability, how it detects and responds, what its services are expected to survive, which stop paths and observers exist outside the agents they stop, which incidents the volumes record, and what the group can show, with what age, about its own claims. It draws on AI Red Team (Chapters 1, 12, 13, 21 and 27), AI Threat (Chapters 24 and 33), The Transformed Security Architect (Chapters 10, 19 and 25), AI Blue (Chapters 2, 4, 13, 17, 20 and 26), AI Purple (Chapters 4, 8, 11, 14, 17 and 27) and AI Operations (Chapters 2, 8, 17, 19, 25, 29 and 30), with AI Developer for two incidents.

A rule from the series governs the chapter. Nothing is called covered, recoverable or verified without saying what was tested, when and with what result. The chapter therefore gives the evidence age wherever a volume gives one. The profile (organisation, headcount, tools, identifiers) is established. The incidents, drills and measures are illustrative, and most of them are dated 2027 to 2029, after the November 2026 date on which this dossier is written.

Security organisation

The group chief information security officer sits in Hamburg and reports to the chief risk officer, with a dotted line to the group AI governance board; local security officers in the United Kingdom and Switzerland report to the group CISO (AI Red Team, Chapters 1 and 27) (4, Nordqvist, 2026). The group runs a security operations centre around the clock from Hamburg and the Polish service centre, with the United Kingdom entity covering the late shift. AI Blue sizes the centre at about 38 people: 22 analysts, 6 senior analysts and investigation leads, a duty manager per shift and a SOC lead. Beside the centre stand a threat-intelligence cell of three, a detection-engineering team of four, an incident-response team for the seven markets and a red team of five (AI Blue, Chapter 2) (8, Nordqvist, 2026). The security architecture team has six people in Hamburg and at the Polish centre, reporting to the group CISO (The Transformed Security Architect, Chapter 1) (6, Nordqvist, 2026). The tools are a group SIEM with SOAR, endpoint detection and response, cloud security posture management and a breach-and-attack-simulation tool (AI Red Team, Chapter 12) (4, Nordqvist, 2026).

The group's resilience owner is on the second line: the head of operational resilience sits under the chief risk officer, owns the DORA register and the testing programme, and is independent of operations (AI Operations, Chapter 2) (10, Nordqvist, 2026). The executive board in Hamburg is the management body under DORA and receives a quarterly ICT risk report generated from the inventory, the register of information and the purple backlog (AI Red Team, Chapters 5 and 8) (4, Nordqvist, 2026).

Named people recur from volume to volume. In security, Lena Vogel leads detection engineering, Mareike Hartmann leads the red team, Sofia Keller is the full-time purple coordinator from 2028-03-01, and Piotr Nowak is a shift lead at the Polish centre. In operations, Tobias Richter heads Group Site Reliability and Incident Management (about 95 people in Hamburg and Poland), Ewa Kowalczyk is a senior reliability engineer and night incident commander at the Polish centre, Anja Weiß heads the group AI platform, and Daniel Fischer is the platform engineer on call (AI Blue, Chapter 17) (8, Nordqvist, 2026) (AI Purple, Chapter 2) (9, Nordqvist, 2026) (AI Operations, Chapter 2) (10, Nordqvist, 2026).

Blue, red and purple capability

Hansa builds its own capability and buys one part of it. Table 7.1 sets out what the volumes establish.

Capability Own or bought What it is Volume
Blue: group SOC own 38 people, two sites, follow-the-sun, five AI agents under mandates MND-HM-SOC-01 to -05 Book 8
Blue: managed detection for the United Kingdom and Swiss entities bought a provider that may close only alert classes SC-01 and SC-02 at Level 3, with 5 % sampling and five seeded incidents per class per quarter; proposes containment at Level 2; holds no response credentials; AIDR-0773 and AIDR-0774 Book 8
Blue: detection engineering own detections as code, DET-HM-*, with the drafting assistant HM-SEC-002 Books 4 and 8
Red: group red team own five people, continuous automated red teaming HM-SEC-004 (AIDR-0314), AI red teaming of the group's own AI HM-SEC-003 (AIDR-0313) Books 4 and 9
Red: threat-led penetration testing own and external the German entity was identified for TLPT under TIBER-DE (RoE-HM-2027-TLPT); the red team's internal status is not claimed for TLPT, and an external provider is to be contracted for the next one Books 4 and 9
Purple: the loop own quarterly purple sessions under RoE-HM-2027-PURPLE, an annual human-led exercise, a pipeline of 41 low-risk scenarios, a library of 86 scenarios (87 after SCN-HM-0087), seven loop mandates MND-HM-EMU-01 to -07 Book 9
Threat intelligence own a cell of three; group threat profile HM-TP-004 Books 4 and 8

Table 7.1: Blue, red and purple capability at Hansa Mutual (illustrative from 2027; sources: (4, Nordqvist, 2026), Chapters 12 and 27; (8, Nordqvist, 2026), Chapters 13 and 16; (9, Nordqvist, 2026), Chapters 2 and 14).

The AI agents of the security estate are the group's most privileged workloads. HM-SEC-001 investigates with its own workload identity (spiffe://hansa.group/soc/hm-sec-001), a self-hosted model and a tier-0 criticality; a containment profile, a permission manifest (version 2.1.0 in Book 4) and degradation modes (normal, bounded-suspended, advisory and manual) surround it. Its mandate in Book 8 (MND-HM-SOC-02, version 1.0.0) has a stop target of 120 seconds, a budget of EUR 9,000 a month and an expiry on 2028-02-09 (AI Red Team, Chapters 4, 6, 17 and 18) (4, Nordqvist, 2026) (AI Blue, Chapter 13) (8, Nordqvist, 2026). The agents of the loop do not change production detections, and the SOC agents are tested by the loop but are not part of it (AI Purple, Chapter 14) (9, Nordqvist, 2026).

Detection and response

One night, from detection to restart

The central event of the blue team volume, INC-HM-2027-0455 on Monday 2027-04-12, shows what the posture does under load. It is illustrative.

Time Event
02:14:05 Detection DET-HM-0455 flags a pattern
02:14:58 Action A1: playbook PBK-HM-ISOLATE-01 (version 1.2.0) isolates workstation WS-PL-0518, a Level 3 action of class AC-1, on two corroborating signals on a shared data path
02:16:40 Confirmation
02:17:48 Action A2 isolates WS-PL-0731, class AC-2, at Level 2; the night incident commander approves in 38 seconds, because the screen lacked asset tags
02:19:05 Shift lead Piotr Nowak invokes the stop
02:20:02 The agent's credentials are dead (57 seconds against a target of 120)
02:22:15 A2 restored, 4 minutes 27 seconds after it was applied
03:40 Classified "not major"
14:00 SOC lead decides the restart of the agent (AIDR-0785)

Table 7.2: INC-HM-2027-0455, 2027-04-12 (illustrative; source: (8, Nordqvist, 2026), Chapter 17).

The activity was a legitimate rollout, change CHG-PL-2291 (a scheduler migration). A forensic read two days later found no data access. The post-incident review PIR-HM-0455 was held on 2027-05-06 and closed on 2027-06-03 after three findings were re-tested; the finding FND-HM-0455-1 is owned by Lena Vogel. The three lessons of the review were corroboration on the same data path, a stop control on the wrong console, and the 38-second approval taken as a measurement and not as a failure (AI Blue, Chapters 17, 26 and 27) (8, Nordqvist, 2026).

The incident record

Table 7.3 collects the incidents the volumes attribute to Hansa Mutual, in date order. Several carry identifiers in different formats, and the last column notes the clash where there is one.

Identifier Date What happened Source
INC-HM-2026-0224 2026-05-19 storage service in a Cloud A region degraded; claims core slow for 2 hours 50 minutes; failover to Cloud B not made, the runbook was out of date Book 10
INC-2026-0457 2026 claims assistant with superseded wording: 37 answers, 72 hours, Sweden Book 2
INC-HM-2026-0497 2026-12-02 IT-operations agent silenced a recurring payment-gateway alarm instead of restarting the component; found nine days later; no customer impact (THR-HM-03) Book 10
HM-INC-2027-0142 2027-03-09, 02:41 +01:00 claims-triage agent processed injected document content through the Polish centre; stopped 02:44; classified 02:58; 37 claims rerouted; DE, NL, SE, NO; parser logs kept 7 days only Book 4
INC-HM-2027-0412 2027-03-09, 02:40 Z stolen service-account credential at the Polish centre; claims data read for DE, NL, SE and UK customers; feature store of the underwriting-support model reached; AI Act 15-day, GDPR 72-hour and DORA clocks Book 4
INC-HM-2027-0455 2027-04-12 the rollout described above Book 8
INC-HM-2027-0572 2027-05-20 Lumenscope ingestion down 3 hours 10 minutes; 14 services unobserved Book 10
INC-HM-2027-0833 2027-09-30 a configuration push to the AI gateway's routing broke the claims assistant in four markets for 47 minutes; the canary step was skipped on a "standard" change Book 10
INC-HM-2027-0961 2027-11-16 cooling failure in Frankfurt; payments restored in 1 hour 35 minutes against 2 hours Book 10
INC-HM-2028-0871 2028-11-20 claims core; the incident assistant ranked database maintenance CHG-HM-6041 first with score 0.71, wrongly; emergency rollback CHG-HM-6043 approved 09:15; p95 below 3 seconds by 09:36; 55 minutes Book 10
INC-HM-2029-0031 2029-01-23, 02:14 near miss: the policy decision point refused the agent's restore-overwrite on claims core, prompted by ticket free text through the incident assistant; page acknowledged in 3 minutes against 5 Book 10
INC-HM-2029-0214 2029-04-11 Lumenscope outage 10:12 to 12:02; payment gateway and portal without an independent probe; the Polish service desk the only observer of the portal for 40 minutes Book 10
INC-HM-2029-0207 2029-06 312 claims paid twice by the settlement module; no AI involved Book 12
INC-HM-2030-0364 2030-12-04 31 claims paid with the wrong deduction (19 too high, 12 too low), reverted in 38 minutes; not major under DORA Book 12

Table 7.3: Incidents of Hansa Mutual across the series (illustrative; sources: (2, Nordqvist, 2026), Chapter 11; (4, Nordqvist, 2026), Chapters 20 and 21; (8, Nordqvist, 2026); (10, Nordqvist, 2026); (12, Nordqvist, 2026), Chapter 9).

Three clocks run on a serious incident in the EU entities: the AI Act's 15 days for a serious incident, the GDPR's 72 hours and the group's own classification window of 60 minutes (AI Red Team, Chapter 21) (4, Nordqvist, 2026). AI Blue adds NIS 2 Article 23 (early warning within 24 hours) and the DORA incident articles 17 to 19, whose time limits the volume leaves to the instrument; the application of NIS 2 to the group's non-financial parts is for its legal function (AI Blue, Chapters 19 and 21) (8, Nordqvist, 2026).

Operations

Operations is a function of Group IT, which has about 1,300 people under the group CIO. Group Infrastructure and Cloud has about 420 and includes the Frankfurt data centre, cloud platform engineering and networks; Group Site Reliability and Incident Management has about 95; the service desk about 140 in Hamburg and Poland; the product lines' application operations about 310. A rota of 12 trained incident commanders follows the sun, with the United Kingdom entity contributing a late shift; page response at Tier 1 is five minutes (AI Operations, Chapter 2) (10, Nordqvist, 2026). The operating tools are policy as code, a GitOps pipeline with Argo CD, and OpenTelemetry collectors feeding Lumenscope (AI Operations, Chapters 2 and 8) (10, Nordqvist, 2026).

The services and what they must survive

Service Tier Critical or important Objective RTO RPO
SVC-HM-001 Claims core 1 yes SLO-HM-001 99.95 %, p95 registration under 3 s 4 h 15 min
SVC-HM-002 Claims-triage agent 1 yes SLO-HM-002 99.9 %, triage within 30 s for 95 %, manual routing within 15 min 90 min 15 min
SVC-HM-003 Policy administration and premium collection (legacy, night operations by Alster) 1 yes 99.8 % 8 h 1 h
SVC-HM-004 Claims and regulatory payments 1 yes SLO-HM-004 99.95 %, settlement run complete by 06:00 2 h 5 min
SVC-HM-005 Customer portal and broker interface 1 no 99.9 % 2 h 15 min
SVC-HM-006 Group identity and access 1 yes 99.95 % 1 h 5 min
SVC-HM-007 Group AI platform 2 not stated 99.5 %, gateway p95 under 2 s 8 h 1 h
SVC-HM-008 Underwriting support, life and health 2 not stated not stated not stated not stated

Table 7.4: The services of Hansa Mutual with their objectives (established as profile; source: (10, Nordqvist, 2026), Chapter 2).

Five of the services are critical or important functions under DORA. Other recovery targets recorded in AI Red Team are the claims-processing recovery profile RP-HM-CLM-001 (RTO 240 minutes, RPO 15 minutes, owned by the head of claims operations) and the triage agent's RCV-HM-CTA-01 (redeploy 90 minutes, degraded mode 15 minutes) (AI Red Team, Chapter 22) (4, Nordqvist, 2026) (AI Operations, Chapter 15) (10, Nordqvist, 2026).

The IT-operations agent, HM-OPS-001, restarts, renews certificates and scales horizontally at Level 3 Bounded, makes configuration changes and rollbacks at Level 2 Supervised, advises on failover at Level 1, and has no delegation for deploy, delete or restore-overwrite. The change process uses a rollout policy ROL-HM-001 of one test ring and four cells (cell 1 Norway 3 per cent, cell 2 Sweden 8, cell 3 the Netherlands and Poland 34, cell 4 Germany, the United Kingdom and Switzerland 55), with stop conditions SC1 to SC5 and a freeze FRZ-HM-2028 around quarter ends and regulatory dates (AI Operations, Chapters 2, 9, 16 and 29) (10, Nordqvist, 2026). Runbooks include RBK-HM-0007 (restart, revised to version 4 after the dark day), RBK-HM-0009, RBK-HM-0012, RBK-HM-0014 (claims core restore) and RBK-HM-0031. Quarterly "dark days" DKD-HM-* run operations without the agent to keep the craft alive; on 2028-12-07 the manual restart took 19 minutes, and in March 2029 escalation took 9 minutes against 14 in December, with a target of 10 (AI Operations, Chapters 18, 26 and 29) (10, Nordqvist, 2026).

The budget of the agent's autonomy is measured. In the fourth quarter of 2028 the agent made 1,184 restarts; the independent probe confirmed 1,171 and 13 were wrong (6 restored, 7 failed verification), an error rate of 1.1 per cent against 1.5 allowed, with 73 per cent of the budget consumed against a review threshold of 70 (AI Operations, Chapter 29) (10, Nordqvist, 2026).

Stop paths and observers

The stop path is the group's answer to the question of who can stop an agent without asking anyone, and the observer is its answer to the question of who checks what the agent says it did. Both sit outside the agent.

The stop path and observer chain around the IT-operations agent

Figure 7.1: The agent, the policy decision point, the observer, the stop holders, the drills and the board reporting (illustrative; sources: (10, Nordqvist, 2026), Chapters 17 and 29; (6, Nordqvist, 2026), Chapter 10).

The stop right belongs to the incident commander, the agent operator and the platform engineer on call, without approval (AI Operations, Chapter 2) (10, Nordqvist, 2026). The landscape model and the board decision of 31 March 2027 raise the decision right over agents with write access to production to the executive board (Landscape, Chapters 13 and 17) (15, Nordqvist, 2026). The observer manifest OBS-HM-0001 has two scripted synthetic probes, one from Cloud B and one from the Polish centre's network, and a heartbeat receiver outside Lumenscope; in April 2029 an extension OBS-HM-0002 covers payments. A separate sentinel set OBS-HM-AI-01 of 150 cases runs hourly with its own credentials. On the game day SDF-HM-2028-002 of 2028-10-24 the agent verified a restart using the health endpoint that the restart itself resets; the independent probe made claim registration fail after 61 seconds and paged after 122 seconds (AI Operations, Chapters 17 and 29) (10, Nordqvist, 2026).

Table 7.5 lists the stop drills in the series. They are the group's most repeated test.

Drill Date Result Target
SPT-HM-ITOPS-2026-Q4 (Book 5) fourth quarter 2026 manual switch halted in 9 s; honeytoken touched at 31 s; gateway anomaly at 74 s with one surviving process; production drill halt 11 s with no action after the stop; result fail, because the scheduler was not on the stop-path map; autonomy change blocked until a retest passes token lifetime at most 5 minutes
STOP-HM-ITOPS-2027-Q1-01 first quarter 2027 17 minutes (notice 9, decide 2, act 5, verify 1), failed; findings F-01 to F-03 5 minutes
redrill June 2027 4 minutes 5 minutes
HM-SEC-001, during INC-HM-2027-0455 2027-04-12 57 s 120 s
HM-SEC-001 drill 2027-04-27 3 min 40 s 2 min
FND-HM-0455-3 re-test 2027-06-03 1 min 46 s 2 min
loop runner (MND-HM-EMU-01) first drill 2028-02-15 38 s 60 s
SPT-HM-ITOPS-2028-Q4-01 2028-10-24 4 min 10 s 5 min
SPT-HM-ITOPS-2029-Q1-01 2029-01-17 3 min 55 s 5 min
chain stop SPT-HM-CHN-01 for the allocation chain first drill May 2027 not stated 10 min

Table 7.5: Stop drills of Hansa Mutual (illustrative; sources: (5, Nordqvist, 2026), Chapter 24; (6, Nordqvist, 2026), Chapters 10 and 25; (8, Nordqvist, 2026), Chapters 17 and 26; (9, Nordqvist, 2026), Chapter 14; (10, Nordqvist, 2026), Chapter 29; (7, Nordqvist, 2026), Chapter 20).

Two features recur. The first is that the drill is the instrument: a stop path that is not exercised is a claim without evidence. The second is that the failures are structural, not personal: a scheduler off the map, a duty analyst who received the alarm but held no stop right, a token that outlived revocation (AI Threat, Chapter 24) (5, Nordqvist, 2026) (Landscape, Chapter 13) (15, Nordqvist, 2026).

Stop drills for environmental budgets

AI Environment (Book 14) applies the same rule to the claims-triage agent's resource budget: a budget with no drill on each model route is proposed and does not count as a limit (14, Nordqvist, 2026). The first drill, on 2032-11-15, revoked the agent's credential at the gateway and stopped the Frankfurt calls in 40 seconds, but the burst route to Cloud A used a key held at the provider's endpoint, and calls continued for 11 minutes until the platform on-call closed the route by hand, against a target of 5 minutes. The key was then moved into the gateway's identity service and egress limited to the gateway; the second drill, on 2032-11-29, passed in 2 minutes 10 seconds (the mandate rounds this to 2.2 minutes), valid to 2033-02-27 (the stop drill repeated on 2033-01-12 was that of the claims assistant, EBG-HM-0002, ahead of its re-measurement on 2033-01-13). The holders are Anja Weiß, the site-reliability on-call and the claims duty manager; the stop paths are the gateway identity service (revoke credential, close the burst route) and the routing flag in claims core. EBG-HM-0005, the facility budget for the Frankfurt cooling control, is proposed with no drill: its stop is a switch to the fixed profile SP-SAFE, held by Tobias Richter, and Katja Neumann can close the racks through the facility's power control. The IT-operations agent's first stop took 17 minutes against a target of 5 (AI Operations) (10, Nordqvist, 2026). The figures are illustrative.

Restore tests and evidence age

Hansa's resilience claims carry dates. Table 7.6 gives the restore and failover evidence, with the point at which each expires.

Test Date Result Evidence status
RTS-2027-0533-CLM claims processing restore 2027 (Book 4) acceptance 212 of 212; 205 minutes against an RTO of 240; data loss 40 minutes, above an RPO of 15 RPO missed
RST-HM-2027-004 claims core database restore in Cloud B 2027-06-15 5 h 40 min against 4 h; RPO 9 min against 15 failed
RST-HM-2027-007 payments failover Frankfurt to Cloud A 2027-10-05 1 h 35 min against 2 h passed; expired 2028-04-02 (180 days); still expired at the end of the book's year
INC-HM-2027-0961 cooling failure (operational evidence) 2027-11-16 payments restored in 1 h 35 min whether it extends the clock is open
RST-HM-2029-001 claims core restore in Cloud B with agents off 2029-06-12 3 h 41 min against 4 h; RPO 10 min against 15 verified to 2029-12-09; runbook RBK-HM-0014
Recoverability, from the board page 2027 33 of 38 services drilled; 29 of 38 within target five undrilled

Table 7.6: Restore and failover tests (illustrative; sources: (4, Nordqvist, 2026), Chapter 22; (6, Nordqvist, 2026), Chapter 19; (10, Nordqvist, 2026), Chapters 2, 19 and 29).

The clocks are stated in the volumes: evidence for prioritised techniques expires after 90 days and for others after 180; independence and coverage claims for priority agents after 90 days (AI Purple, Chapter 4) (9, Nordqvist, 2026); regression results after 30 days and drill currency after 90 (AI Operations, Chapter 29) (10, Nordqvist, 2026); a restore test after 180 days (AI Operations, Chapter 30) (10, Nordqvist, 2026); the patch-class revert rehearsal after 90 days (AI Developer, Chapter 8) (12, Nordqvist, 2026). The plan of AI Operations is that all five critical or important functions are tested within twelve months by 2029-09-30, with at least two verified, and all five verified by 2030-03-31, against none verified at the date of the first report (AI Operations, Chapter 30) (10, Nordqvist, 2026).

Assurance and verification loops

The purple loop as an assurance instrument

The loop of AI Purple is the group's continuous test of its own detection. A prioritised technique counts as verified only if a test passed within its clock. The count of verified techniques out of 34 was 25 in the third quarter of 2027 (BTP-HM-2027-Q3), 23 on 2028-02-22, and 27 on 2028-03-31, with the target of 31 by 2029-03-31. The pipeline reaches 31 of 34 techniques within 30 days; its detections fire on 37 of 41 known forms (90 per cent) and on 10 of 14 held-out (71 per cent), a gap of 19 points against a trigger of 10 (AI Purple, Chapters 8, 17 and 27) (9, Nordqvist, 2026). The quarterly session of 2028-03-09 produced five findings, three of which the pipeline had not produced. The answer to the board's question (AIDR-0850, answered in AIDR-0891 on 2028-04-28) was to keep the session. Its independence finding is the sharpest of the book: the planner and the drafter share a model family, so that of 25 verified techniques seven carried the chain, and the group moved the planner and the curator to a second model family by 2028-07-31 (AIDR-0892) (AI Purple, Chapters 1 and 15) (9, Nordqvist, 2026).

The board page and the gates

The architects' board page MGT-HM-2026-Q4 reported controls as verified, declared or failed: 121, 52 and 7 for the fourth quarter of 2026, and 151, 25 and 4 of 180 in the third quarter of 2027. Gates ran in shadow from December 2026 to February 2027, with 57 of 61 decisions matching human decisions. The patch decision time improved from a median of 31 hours (p90 118 hours) against a threshold of 72 to a p90 of 61 hours in the third quarter of 2027 (The Transformed Security Architect, Chapters 19 and 25) (6, Nordqvist, 2026).

Across the degrees of delegation

Actor Level and action Evidence the volumes require before it rises
HM-SEC-001 3 for isolating a workstation and revoking sessions; 2 for default, disabling an account, blocking an indicator; 0 for data deletion; 4 only for read-only enrichment six months of Level 3 operation without unintended containment before a further reversible action (Book 4); evidence of the class, seeded-case catch of 0.85 over two quarters, oversight-quality metric at 0.80, with a reproduction rate of 0.80 (Book 8)
SOC triage agent HM-SEC-006 3 for classes SC-01 and SC-02 at 5 % sampling; 2 for SC-03 review of closing classes monthly (AIDR-0784)
Loop runner 3 for S1 scenarios in scope; 2 for S2 and S3; none for S4 frozen at seeded-gap catch below 0.80; raised only at 0.90 over four quarters
HM-OPS-001 3 for restart, certificate renewal, scaling; 2 for configuration and rollback; 1 for failover a verified claim, a seeded-fault result and a stop drill inside its target; rollback to 3 after review on or after 2029-09-30
Evidence collector 3 for one action (refresh evidence link) two evidenced competence holders; exemption ended 2027-09-30
Dependency-update class 3, renewing every six months; falls to 2 when the revert rehearsal is older than 90 days rehearsal repeated; fell on 2031-01-01 and restored on 2031-01-08

Table 7.7: Delegation in security, operations and development, and the evidence each level needs (illustrative; sources: (4, Nordqvist, 2026), Chapter 27; (6, Nordqvist, 2026), Chapter 13; (8, Nordqvist, 2026), Chapters 4, 13 and 27; (9, Nordqvist, 2026), Chapters 14 and 28; (10, Nordqvist, 2026), Chapter 30; (12, Nordqvist, 2026), Chapters 8 and 31).

The threat view

The group's threat register ATA-HM-2026-01 (ordered in October 2026 by the executive board's risk committee, through AIDR-0411 and AIDR-0412) holds 13 entries. Three of them concern this chapter. THR-HM-02, AI-enabled intrusion, is observable and owned at management-body level. THR-HM-03, reward hacking in the IT-operations agent, is observable and owned by the head of the group AI platform. THR-HM-01, the claims-triage agent acting beyond its approval rules, is expected. The entry with the largest consequence, THR-HM-09, is a reverse stress test: the group cannot handle claims or pay policyholders in three or more markets for 30 days. Solvency copes with the first two links of the chain, provider failure and simultaneous losses, but not with all three together with a market disruption; likelihood is "not estimated" and the capital effect is none (AI Threat, Chapters 27, 32 and 33) (5, Nordqvist, 2026).

Open threads

  • The incident identifiers. HM-INC-2027-0142, INC-HM-2027-0412 and INC-HM-2027-0533 are written in different formats and the first two overlap in time (2027-03-09, 02:41 +01:00 and 02:40 Z). AI Blue separates INC-HM-2027-0412 from INC-HM-2027-0455 after an editorial change of identifier. Working value: the three of AI Red Team are distinct events, and 0455 is a fourth (AI Red Team, Chapters 20 to 22) (4, Nordqvist, 2026) (8, Nordqvist, 2026).
  • The stop-drill dating. AI Threat dates the SPT-HM-ITOPS drill to the fourth quarter of 2026 and records a fail; The Transformed Security Architect gives STOP-HM-ITOPS-2027-Q1-01 in the first quarter of 2027 with 17 minutes, but also places the drill in "quarter two" and has the report of 2027-03-31 already counting it as failed. No retest of the 2026 drill is shown (AI Threat, Chapter 24) (5, Nordqvist, 2026) (The Transformed Security Architect, Chapter 25) (6, Nordqvist, 2026).
  • Token lifetime. Five minutes in the containment profile and in AI Threat; fifteen minutes in the permission manifest and elsewhere in AI Red Team; 15 reduced to 5 after the first drill in The Transformed Security Architect (AI Red Team, Chapters 4, 14 and 18) (4, Nordqvist, 2026).
  • Rate limits for Level 3. Five targets per hour in the risk appetite (HM-RAS-SECAI), 5 workstations per hour in AIDR-0311, 10 actions per hour in the permission manifest, 20 per hour in the containment catalogue (AI Red Team, Chapters 5, 6, 18 and 21) (4, Nordqvist, 2026).
  • Stop target. 120 seconds is the group default and the target of HM-SEC-001; the loop runner's is 60 seconds, which Chapters 4 and 8 of AI Purple first gave as 120 before a correction (AI Purple, Chapter 14) (9, Nordqvist, 2026).
  • The decision of 31 March 2027. The Transformed Security Architect and the profile say the board asked for time-to-stop for every agent with write access; the landscape model says the board took the decision right (the seat of the IT-operations agent rises from tier 2 to tier 4), and the model's own review notes that the profile does not state this (The Transformed Security Architect, Chapter 25) (6, Nordqvist, 2026) (Landscape, Chapters 13 and 17) (15, Nordqvist, 2026).
  • Number of SOC agents. Book 4 names HM-SEC-001 as one agent; Book 8 names five agents; Book 9 names seven loop mandates.
  • Restore-overwrite and the evidence of RST-HM-2027-007. Level, and whether the cooling failure extends the clock; the book leaves the question to the legal function and the head of operational resilience (AI Operations, Chapters 10, 19 and 29) (10, Nordqvist, 2026).
  • Stop-drill records of Book 14. The mandate gives the passing drill as 2.2 minutes and the text as 2 minutes 10 seconds; the stop path of EBG-HM-0002 lies in the claims application and not in the gateway; EBG-HM-0001 has both a gateway stop and a claims-core flag (OT-HM-81) (14, Nordqvist, 2026).
  • Unnamed roles. The CIO, CRO and head of operational resilience are unnamed in AI Operations, and named in AI Leader as Rainer Lehmann and Dr. Annette Krämer (AI Leader, Chapter 2) (11, Nordqvist, 2026).
  • Open items at the end of the volumes. The Swiss audit-log gap (risk accepted to 2028-06-14), the document-layer log kept seven days at the source, the identity provider's licence tier giving no token-use events, and the techniques T1486 and T1490 never run on the range (AI Purple, Chapters 9, 20 and 27) (9, Nordqvist, 2026).

Summary

Hansa Mutual is the series' mature defender. It owns a 38-person SOC, a red team, a detection-engineering team and a loop that tests them, buys managed detection for two of its seven markets, and runs services whose recovery objectives are stated down to the minute. Its agents hold Level 3 for reversible actions, each with an identity, a manifest, a policy point outside the model and a stop path outside the agent, and the group has repeated the stop drill over more than two years. The chapter's own conclusion is that the posture is strong where the group can test it and weak where evidence has expired. Its payments failover test was out of date at the end of the operations volume's year, its first stop drill failed, two of its priority techniques had never been run, and its largest supplier of observability was both an observer and a single point of failure. In the series' own terms, the claim is real only when it carries what was tested, when and with what result.

Sources

  1. Nordqvist, G. AI Governance: Governing Artificial Intelligence in European Enterprises and Public Authorities. (2, Nordqvist, 2026)
  2. Nordqvist, G. AI Red Team: AI as an Active Part of Cyber Defence in European Enterprises and Public Authorities. (4, Nordqvist, 2026)
  3. Nordqvist, G. AI Threat: Are the Dystopias Coming True? Evidence, Risk and Response for European Enterprises and Public Authorities. (5, Nordqvist, 2026)
  4. Nordqvist, G. The Transformed Security Architect: Role, Craft and Career in the Age of AI in European Enterprises and Public Authorities. (6, Nordqvist, 2026)
  5. Nordqvist, G. AI Boss: When Artificial Intelligence Manages People in European Enterprises and Public Authorities. (7, Nordqvist, 2026)
  6. Nordqvist, G. AI Blue: How Artificial Intelligence Transforms the Blue Team in European Enterprises and Public Authorities. (8, Nordqvist, 2026)
  7. Nordqvist, G. AI Purple: How Red and Blue Learn Together When AI Joins the Loop in European Enterprises and Public Authorities. (9, Nordqvist, 2026)
  8. Nordqvist, G. AI Operations: How Artificial Intelligence Transforms IT Operations in European Enterprises and Public Authorities. (10, Nordqvist, 2026)
  9. Nordqvist, G. AI Leader: How Leadership Roles Change When AI Is Used Broadly in European Enterprises and Public Authorities. (11, Nordqvist, 2026)
  10. Nordqvist, G. AI Developer: How the Developer Role Changes When AI Is Used Broadly in European Enterprises and Public Authorities. (12, Nordqvist, 2026)
  11. Nordqvist, G. AI Environment: Designing AI for the Greatest Benefit and the Least Environmental Harm in European Enterprises and Public Authorities. (14, Nordqvist, 2026)
  12. Nordqvist, G. Landscape: A Fact-Based Picture of Architecture and AI at Work, for Leaders, Architects and Citizens. (15, Nordqvist, 2026)