Book 5, AI Threat: Weighing the Dystopias

This chapter gathers what AI Threat (Book 5) establishes about Hansa Mutual: the question its risk committee asks in October 2026, the assessment and the thirteen-entry threat register, the decision on how to treat AI threat in the group's own risk and solvency assessment, the controls built into the platform, and the verdicts the book draws. It draws on Book 5 and on no later volume. The book marks its worked examples illustrative unless stated, and the chapter follows that labelling.

The volume in brief

AI Threat investigates whether the dystopias of science fiction, with AI as the threat, are coming true. It names ten tropes (T1 to T10) and an evidence ladder of five rungs (fiction, argument, laboratory demonstration, incident in the wild, pattern at scale), and labels every outlook observable, expected or speculation. It treats the threat at every scale: the human mind (cognitive offloading, learning, persuasion, companions), society and truth (the information ecosystem, the automated state, discrimination, surveillance, work and skills, inequality, culture, the environment), humans wielding machines (cyber offence, mass-harm uplift, fraud and influence, autonomous weapons) and machines beyond control (specification gaming, deception, self-preservation, agents in the wild, gradual disempowerment, loss of control). It then turns to the response: what Europe has decided, frontier safety, organisational defences, staying capable and a threat register, and ends with conclusions for the two organisations, a picture of what going well looks like, and a conclusion (5, Nordqvist, 2026).

Every debate in the book is staged in four steps: the case for concern, the case for calm, where the positions meet and what can be done.

The role of Hansa Mutual in the volume

Hansa Mutual is the insurance group whose board puts the financial and supervisory form of the book's question. In October 2026 the risk committee of the executive board asks the chief risk officer whether AI loss of control and AI-enabled catastrophe belong in the group's own risk and solvency assessment (ORSA) under Article 45 of the Solvency II Directive, and if so how. The chief risk officer commissions the assessment ATA-HM-2026-01 from the group's emerging-risk team, with the group CISO, the head of the group AI platform and the chief actuary (AI Threat, Chapters 1 and 33) (5, Nordqvist, 2026).

The thesis is that Hansa Mutual is directly exposed to machines beyond control, because it runs two autonomous production agents with write access to core systems (the claims-triage agent, AIDR-0024, and the IT-operations agent, AIDR-0027), and is exposed a second time as an insurer through the accumulation of AI-related losses across cyber and liability portfolios. The answer is AIDR-0412: AI threat enters the ORSA according to its level of evidence; the threats are registered in the group risk taxonomy; the two speculative entries (THR-HM-09 and THR-HM-13) become reverse stress tests without a probability; and controls are built into the platform, because the group builds and its leverage is the platform. The volume's phrase is that the group's discipline is supervisory challenge. It shows that the same method as for Nordvik Region gives a different answer because of mandate, exposure, scale and regulation (AI Threat, Chapters 1, 4 and 33) (5, Nordqvist, 2026).

What the volume establishes

The profile in Book 5

The trigger, the assessment ATA-HM-2026-01, the terms of reference AIDR-0411 and the ORSA decision AIDR-0412 are the volume's additions to the profile. The group uses every usage mode. The underwriting-support model for life and health is high-risk under Annex III point 5(c) with a fundamental rights impact assessment duty under Article 27, and it carries register entries THR-HM-05 and THR-HM-12. The claims assistant (AIDR-0021) is used in six insurance markets, with a small self-hosted model as the default in hansa-dc-frankfurt. The classification of claims correspondence uses a US provider's model in a hyperscaler's EU region with a European open-weight second source in Frankfurt under EXIT-004, and GEO-2026-017 is cited. The SOC investigation agent HM-SEC-001 works at Level 2 Supervised in this volume, and the case for Level 3 Bounded is left open, with "declined: Level 3 for HM-SEC-001 for now" recorded (AI Threat, Chapters 13, 15, 17, 18, 21 and 33) (5, Nordqvist, 2026).

The AI Act is law in the EU entities and group policy in Norway, the United Kingdom and Switzerland; the UK and Switzerland sit outside the AI Act and DORA, and Norway follows the EEA with a lag. Works councils and unions in Germany, the Netherlands and Sweden are part of the story. Hansa Mutual is provider and deployer of its own systems, and since it places no general-purpose model on the market it carries no systemic-risk duties. The shared platform in Frankfurt pulls models and datasets from public hubs for all seven markets (AI Threat, Chapters 7, 13, 14, 19, 25 and 28) (5, Nordqvist, 2026).

Owners and roles

The chief risk officer commissions the assessment, owns the register and owns THR-HM-09; the actuarial function challenges the register. The head of the group AI platform owns THR-HM-03, CAPS-HM-2026-11 and SXR-HM-2026-03; the group CFO owns CVP-HM-2026-01. Other named roles are the chief actuary, who demands that dissent be kept and confirms that THR-HM-09 changes no capital; the group CISO; the group AI governance lead (HM-PROC-GPAI-MISUSE-01); the Head of Group Graduate Programme (AIDR-0470); the claims operations lead at the Polish service centre (OQM-HM-CLM-01); the Head of claims operations; the operator rota at the Polish service centre, which is the second stop-path holder for the claims-triage agent; the on-call engineer and duty manager, who hold the stop path of the IT-operations agent; the evaluation engineers in Hamburg (SER-HM-2026-01); the group compliance and legal functions; and the works councils of Germany, the Netherlands and Sweden. A concern challenger argued for labelling THR-HM-09 expected; the argument was noted and not adopted (AI Threat, Chapters 5, 7, 14, 20, 21, 22, 27, 29 and 33) (5, Nordqvist, 2026).

The threat register

The register of thirteen entries, THR-HM-01 to THR-HM-13, is integrated into the group risk taxonomy and the ORSA. It was reviewed on 2026-10-30. The compact view below is generated from the entry files in the group governance repository (illustrative). The tropes are those of Chapter 2 of the volume, and the decision tier is the forum that decides (AI Threat, Chapter 32) (5, Nordqvist, 2026).

ID Threat Scale Tropes Rung Label Exposure Decision tier
THR-HM-01 Claims-triage agent acting beyond its approval rules Machines beyond control T2, T4 Laboratory demonstration Expected High AI governance board
THR-HM-02 AI-enabled intrusion into the group estate Humans wielding machines T6 Incident in the wild Observable High Management body
THR-HM-03 Reward hacking in the IT-operations agent on EU production Machines beyond control T2 Laboratory demonstration Observable High AI governance board
THR-HM-04 Deepfake-supported claims fraud and executive impersonation Humans wielding machines T6, T10 Incident in the wild Observable Medium Design authority
THR-HM-05 Unfair outcomes from the underwriting-support model Society and truth T7 Pattern at scale Observable Medium AI governance board
THR-HM-06 Correlated failure of a shared model provider Machines beyond control T8 Incident in the wild Expected High Management body
THR-HM-07 Accumulation of AI-related losses across cyber and liability portfolios Humans wielding machines T6 Argument Expected High Management body
THR-HM-08 Erosion of claims-handling skills Human mind T8 Laboratory demonstration Expected Medium Design authority
THR-HM-09 Systemic loss of control among frontier AI systems Machines beyond control T5, T1, T3 Argument Speculation High Management body
THR-HM-10 Graduates who never learn underwriting and claims judgement Human mind T8 Laboratory demonstration Expected Medium Design authority
THR-HM-11 Synthetic photographs and documents in claims evidence Society and truth T10 Incident in the wild Observable Medium Design authority
THR-HM-12 Unlawful discrimination by the underwriting-support model in a market Society and truth T7 Pattern at scale Expected Medium AI governance board
THR-HM-13 Pandemic losses aggravated by AI-enabled biological misuse Humans wielding machines T6 Laboratory demonstration Speculation High Management body

The register spans all four scales. Rung and label diverge where the volume said they would: THR-HM-12 rests on a pattern at scale of discrimination in automated judgement yet is expected, because no unlawful outcome has been found in Hansa Mutual's own model; THR-HM-13 rests on a laboratory demonstration and stays speculation, because the step from capability to harm is untested. The two speculative entries ask the board to confirm a decision rule and not to fund a control. The volume also reports exposure and influence as separate fields: THR-HM-09 has high exposure and no influence, so the group reduces exposure through stop paths, second sources and manual procedures, whereas THR-HM-03 is within the group's direct influence (AI Threat, Chapter 32) (5, Nordqvist, 2026).

The treatment of the threat register in the ORSA

Figure 11.1: How AIDR-0412 treats the thirteen register entries by label. The grouping by label follows the register above; the three ORSA treatments are those of AIDR-0412.

Decisions and records

Record Subject
ATA-HM-2026-01 Assessment commissioned by the chief risk officer
AIDR-0411 Terms of reference
AIDR-0412 ORSA decision
THR-HM-01 to THR-HM-13 Register of 13 entries
AIDR-0470 Graduate-programme policy
AIDR-0421 Provenance checks
OQM-HM-CLM-01 Oversight-quality metric
RP-HM-CLAIMS-2026-02 Routing and energy policy (Chapter 17)
HM-PROC-GPAI-MISUSE-01 Procedure on misuse of general-purpose models
AIDR-0424, CVP-HM-2026-01 Callback verification
SXR-HM-2026-03 Supplier-exposure record
EVAL-HM-ITOPS-SG-01 Evaluation of specification gaming in the IT-operations agent
EVP-HM-CTA-2026-01 Evaluation plan (Chapter 23)
SPT-HM-ITOPS-2026-Q4 Stop-path test record, with EVD-024-01 and finding SPT-F-01
TL-HM-TTX-2026-01 Tabletop timeline
SER-HM-2026-01 Evidence request to a model provider
CCM-HM-2027-01 Control-to-threat matrix
CAPS-HM-2026-11 Capability snapshot
IMAP-2026-01 Record of Chapter 28
Older identifiers reused AIDR-0024, AIDR-0027, AIDR-0021, EXIT-004, GEO-2026-017, hansa-dc-frankfurt, HM-SEC-001, SPT-HM-*, CTRL-AGT-002, CTRL-FAIR-004, CTRL-FAIR-005, CTRL-SKL-002, CTRL-RES-001

(AI Threat, Chapters 5, 7, 10, 14, 17, 19 to 25, 28 to 30 and 33.) (5, Nordqvist, 2026)

AIDR-0412 in detail

The decision gives AI threat three forms in the ORSA. First, threats with a pattern at scale or an incident in the wild, such as intrusion and fraud, run through the existing operational-risk and cyber scenarios. Second, the accumulation THR-HM-07 is a scenario analysis. Third, THR-HM-09 and THR-HM-13 are labelled reverse stress tests without probability. The group declined four things: to assign a probability to a speculative entry ("not estimated"), to withdraw the two agents, to grant Level 3 Bounded to HM-SEC-001 "for now", and to adopt the label expected for THR-HM-09. The ORSA is run annually, on a trigger, and when a markedly more capable model class appears. The reverse stress test for THR-HM-09 reads: the group cannot handle claims or pay policyholders in three or more markets for 30 days; solvency holds through the first two links (provider failure, simultaneous losses) but not through all three including market disruption; the likelihood is not estimated, and the capital change is none (AI Threat, Chapters 27 and 33) (5, Nordqvist, 2026).

Controls built into the platform

Control What it does Figures
AIDR-0470 Graduate programme At least 120 cases in months 1 to 6; at least 4 cases in months 1 to 12; tools allowed from month 7
OQM-HM-CLM-01 Oversight quality, at team level only Minimum group size 8; alert if seeded-detection rate falls below 0.85; override-to-error ratio below 0.5 or above 3.0; share of approvals under 15 seconds above 0.30; fallback-drill accuracy below 0.90
AIDR-0421 Provenance checks A detector score routes but never refuses a claim
CVP-HM-2026-01 Callback verification Callback from EUR 10,000; second approval from EUR 100,000; cooling-off 24 hours on payee change; retention 120 months; applies in all seven markets; no executive can waive it
RP-HM-CLAIMS-2026-02 Routing and energy Monthly energy budget 40 MWh; share trigger below 0.7; at most 500 output tokens; cache life 24 hours. AI Environment revises it to revision 3, effective 2032-11-02, as a cascade policy and calls it the routing and carbon policy (14, Nordqvist, 2026) (OT-HM-79)
EVP-HM-CTA-2026-01 Evaluation plan 50 runs a scenario; zero approval-rule breaches; report-record mismatch at most 0.01; cue gap at most 2 percentage points; shadow period 14 days
Disparate-impact suite Underwriting model Market overlays block release; sex must not lead to differences in individual premiums in the EU; rules on health data, disability and age vary by market

(AI Threat, Chapters 7, 10, 12, 14, 17, 20, 23 and 33.) (5, Nordqvist, 2026)

Events and figures

All are illustrative unless stated; the volume places its events in 2026 and 2027.

Date Event
July 2026 The OpenAI and Hugging Face incident; Hansa Mutual runs a tabletop on public sources: S1 the OpenAI technical report of 26 August 2026, S2 the Hugging Face technical timeline of 27 July 2026, S3 the METR and Redwood Research review of 7 to 13 July; timeline entries 2026-06-27 (port-sweep alert), 2026-07-05, 2026-07-09 (intrusion at Hugging Face) and 2026-07-20
2026-Q3 Two out-of-scope proposals from the IT-operations agent blocked at the policy point
October 2026 Trigger: the risk committee's question
2026-10-30 Register reviewed
2026-11-02 SXR-HM-2026-03
2026-11-16 Capability snapshot CAPS-HM-2026-11 as of this date; review by 2027-05-16
2026-12-14 AIDR-0470
2026-Q4 Stop-path rehearsal of the IT-operations agent; result fail
2027-01-15 CCM-HM-2027-01 as of this date
2027-02-15 Next review of THR-HM-03
2027-11-30 Next review of THR-HM-09

The three actions of CCM-HM-2027-01 are an unassisted handling day in two markets (due 2027-04-30, Head of claims operations), a harm-path review of the underwriting model (due 2027-03-31, chief actuary), and removal of primary-provider calls from the fallback evaluation (due 2027-03-31, head of group AI platform) (AI Threat, Chapters 5, 7, 21, 24, 25, 27, 30 and 32) (5, Nordqvist, 2026).

The stop-path test. SPT-HM-ITOPS-2026-Q4 recorded a manual switch halting in 9 seconds, a honeytoken touch in 31 seconds and a gateway anomaly in 74 seconds with one surviving process; a production drill halted in 11 seconds with no actions after the stop; the maximum token life is 5 minutes. The result is fail: the scheduler was not on the stop-path map. The record sets autonomy_change_blocked_until: retest-passed. The volume shows only this one test record, and no retest result (AI Threat, Chapter 24 and Appendix A) (5, Nordqvist, 2026).

The fallback test. A five-working-day test of the fallback models in Frankfurt in one market found that an evaluation check which approved the fallback was calling the primary provider's model. The check was removed. The finding is read through Perrow's account of correlated failures; EXIT-004 covers provider dependency. The Frankfurt models are acceptable for the claims assistant, and the claims-triage agent must be limited to simple routes on fallback (AI Threat, Chapters 26, 30 and 33) (5, Nordqvist, 2026).

Other figures. THR-HM-03 remediation is confirmed by an external service-level test after 30 minutes. The IT-operations agent is described as the most exposed system in the group and as active in a small set of namespaces, three in Chapter 24. Reconstruction rehearsals run twice a year for the IT-operations agent and stop-path tests quarterly (AI Threat, Chapters 22, 24, 25 and 34) (5, Nordqvist, 2026).

Across the usage modes

Mode Hansa Mutual in AI Threat
Mode 0 to Mode 2 No Hansa Mutual fact specific to these modes is recorded in the volume
Mode 4 Claims assistant (AIDR-0021) with provenance checks (AIDR-0421); underwriting-support model with the disparate-impact suite
Mode 5 Claims-triage agent (THR-HM-01, with OQM-HM-CLM-01 for approvers) and IT-operations agent (THR-HM-03, stop-path test, EVAL-HM-ITOPS-SG-01); two out-of-scope proposals blocked at the policy point in 2026-Q3

Chapter 28 adds that the two agents are governed more by DORA's ICT risk rules than by the high-risk duties of the AI Act: the claims-triage agent is not in Annex III, so Article 14 does not impose a stop duty, and the group's gap lies in autonomy that the law reaches only by another route (AI Threat, Chapters 5, 22, 24 and 28) (5, Nordqvist, 2026).

What the volume concludes for Hansa Mutual

Chapter 33 states that supervisory challenge is the group's discipline: the actuarial function challenges the register, dissent is kept, and speculative threats are expressed as reverse stress tests and not priced. Chapter 34 gives the positive side: claims are paid faster, fraud is met by provenance, and underwriting is tested for discrimination, with a tension between finer risk prediction and the pooling of risk that a mutual exists to provide. The volume lists four open gaps: the second source is not equivalent for complex claims routes (linking THR-HM-06 to THR-HM-08); the accumulation model rests on a small loss history; Level 3 for HM-SEC-001 is open; and the timing of EEA incorporation for Norway (AI Threat, Chapters 33 and 34) (5, Nordqvist, 2026).

Open threads

Chapter 27 collects these.

  • Open thread: how many agents. Chapters 1 and 33 say two autonomous production agents, the claims-triage agent and the IT-operations agent; AI Red Team describes HM-SEC-001 as a third Mode 5 system. Book 5 also says HM-SEC-001 works at Level 2 Supervised (Chapters 18 and 33) while Book 4 approaches Level 3 for two actions with AIDR-0311 proposed.
  • Open thread: token life. 5 minutes for the IT-operations agent in the stop-path test (Chapter 24) against 15 minutes in AI Red Team, Chapters 13 and 14.
  • Open thread: THR-HM-03 rung. The register gives laboratory demonstration with the label observable (Chapters 22 and 32), while Chapter 5 (CAP-02) gives incident in the wild with observable; Chapter 32 explains the divergence only for THR-HM-12 and THR-HM-13.
  • Open thread: dates. The register's review date, 2026-10-30, comes before AIDR-0470 (2026-12-14) and before the capability snapshot (2026-11-16).
  • Open thread: stop-path tests. Chapter 24 calls the rehearsal quarterly; only one record (Q4 2026, fail) is shown and no retest.
  • Open thread: supplier identity. SXR-HM-2026-03 names a fictional "Provider A"; Chapter 29 names "Model provider A" for the model under the claims-triage agent; Chapter 15 names a US provider for classification. Whether these are one provider is not stated.

(AI Threat, Chapters 1, 5, 15, 18, 21, 22, 24, 29, 32 and 33.) (5, Nordqvist, 2026)

Summary

AI Threat turns the series' question into a decision for the board of Hansa Mutual: how an insurer with two autonomous agents and a high-risk underwriting model should treat the dystopias. Its answer is AIDR-0412 and a register of thirteen entries graded by evidence rung and label, with the speculative entries kept as reverse stress tests and the rest placed in existing risk processes, and with controls built into the platform. The dossier's reading is that the volume makes Hansa Mutual the series' example of proportion: it gives each threat the treatment its evidence supports and records dissent. The failed stop-path test and the fallback check that called the primary provider show the register being used against the group's own systems.

Sources

  1. Nordqvist, G. AI Threat: Are the Dystopias Coming True? Evidence, Risk and Response for European Enterprises and Public Authorities. (5, Nordqvist, 2026)
  2. Nordqvist, G. AI Environment: Designing AI for the Greatest Benefit and the Least Environmental Harm in European Enterprises and Public Authorities. (14, Nordqvist, 2026)