# The Next Cybersecurity Crisis Is a Crisis of Sovereignty

> Pearce Robinson, sovereign systems designer, data scientist and creator of the Belel Protocol, examines cyberattacks and autonomous-agent failures, and sets out his architecture for national control of data, compute, models and execution authority under the Concordium Mandate.

By Pearce Robinson · PearceRobinson.com · Investigation · Published 2026-11-15
Canonical URL: https://pearcerobinson.com/reporting/next-cybersecurity-crisis-sovereignty-belel-protocol
Topics: Belel Protocol, Concordium Mandate, Pearce Robinson sovereign systems designer, Pearce Robinson data scientist, AI security, AI safety, AI governance, AI alignment, Constitutional AI, Claude constitution, Anthropic, autonomous AI agents, AI agent sandbox escape, sovereign systems design, sovereign AI, digital sovereignty, critical infrastructure cybersecurity, data sovereignty, model weights escrow, runtime enforcement, AI sovereignty Caribbean, data centres Trinidad, rogue AI cybersecurity, Amadeus breach airlines, operational sovereignty, supermarket freezer cyberattack Europe, FAA nationwide ground stop, sovereign AI Oman

## Key evidence — source-record extracts

- One provider, many airlines, many jurisdictions: Japan Airlines reported about 92,000 affected customers and Singapore Airlines about 46,000 booking records; Cathay Pacific, Qantas and Thai Airways reported actual or potential exposure. Records included passport details, contact details and itineraries. No compromise of flight-control systems was reported. [Read the source record](https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record#amadeus)

- The American deregulatory sequence: Executive Order 14179 (January 2025), the AI Action Plan (July 2025) and the June 2026 order on advanced AI innovation and security prioritise reduced regulatory barriers while acknowledging security. The United States has not abandoned all AI oversight. [Read the source record](https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record#us-policy)

- Critical infrastructure needs AI-specific risk management: Recognises the need for AI-specific risk management where AI is used in critical infrastructure. [Read the source record](https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record#nist)

Stolen personal information can remain in circulation long after the systems breached in a cyberattack have been restored. ¶1

An airline can resume operations, a bank can reopen its digital services and a ministry can recover from an outage. Once sensitive personal information has been extracted, copied and distributed beyond its custodian's control, retrieving every copy may be impossible. ¶2

Restoring a service does not restore control over stolen information. Artificial intelligence enables criminals to combine and analyse that information while governments give automated systems more access to essential services. Governments need to establish who controls those systems, whose laws govern them and how they can intervene when a system acts against the country's interests. ¶3

## I. One provider, many countries

In late October and early November, airlines on several continents disclosed exposure to a cybersecurity incident involving Amadeus, a major travel technology provider. Japan Airlines reported about 92,000 affected customers; Singapore Airlines identified around 46,000 potentially affected booking records. Cathay Pacific, Qantas and Thai Airways reported actual or potential exposure. Some records included passport information, contact details and itineraries.[1] ¶5

None of this involved the compromise of aircraft flight-control systems. An attack on one shared provider can expose passenger records held for many airlines across several jurisdictions. Dependence on that provider concentrates the risk. ¶6

## II. A fortnight that should end the complacency

In mid-October, a cyberattack disrupted as many as 700 grocery stores in France. Three weeks later, on 2 November, supermarkets in at least ten European countries — Belgium, Estonia, Finland, Germany, Ireland, Italy, Luxembourg, the Netherlands, Portugal and Sweden — found that someone had remotely changed the temperature settings on their freezers and cold rooms. Units were switched off or forced into defrost; some reportedly reached 50C.[12] More than 115 stores were hit, among them about 40 Ahold Delhaize sites, 15 Carrefour and 15 Intermarché stores in Belgium, 13 SuperValu stores in Ireland, and ICA and Coop branches in Sweden.[13] The European Union's food commissioner, Christophe Hansen, called it 'a deliberate attempt to disrupt our food supply chain'.[25] France has since attributed its own supermarket attacks to Shahid Kaveh, a unit it links to Iran's Islamic Revolutionary Guard Corps. Iran calls the accusation 'baseless and hostile', and it is not known whether the wider European wave is connected.[14] ¶8

Technically, this is the attack surface I worry about most: operational technology connected to the internet for the convenience of remote monitoring. A refrigeration controller is a closed feedback loop around a setpoint. Change the setpoint and the plant obeys perfectly — nothing is stolen, nothing is encrypted, and the damage is physical, measured in degrees and tonnes of food. A fleet-wide setpoint change at four in the morning across unrelated chains is a textbook multivariate outlier; any properly trained baseline model would flag it within a sampling interval. The problem is architectural. When the control channel and the monitoring channel run through the same vendor cloud, the system that should raise the alarm is the system that was compromised. ¶9

Then, on 13 November, the United States grounded every commercial flight in the country. According to the Federal Aviation Administration, legacy equipment became unstable when data flooded its telecommunications network at about 1,000 connection attempts every four to five seconds, traced to two sites where the contractor L3Harris had been deploying a new router.[15] Seventeen of 22 air route traffic control centres reported instability; the stop lasted more than four hours, and engineers needed 45 minutes to find the source because maintenance was logged by hand, on paper.[16] Officials say there is no evidence of a cyberattack and have ruled out any role for SMART, the FAA's new AI planning tool.[17] The FBI is nonetheless assisting, and the cause has not been formally closed.[18] ¶10

It was, on current evidence, an accident: one misbehaving router on a contractor-operated network produced a broadcast storm large enough to halt the airspace of the world's largest aviation market. An adversary or an autonomous agent with network access could study the same failure. Preventing its spread requires network monitoring, segmentation, a live inventory of changes and clear authority to isolate the affected equipment. ¶11

## III. Possession is not control

On 11 June 2026 I led a lecture on Sovereign AI Management for the Omani government, convened with the Ministry of Transport, Communications and Information Technology, the AI Studio in Muscat and AWJ.[11] Its purpose was prevention: to give a state the architecture to avoid precisely the failures this investigation documents, weeks and months before the AISI incident, the supermarket sabotage and the American ground stop made them headlines. I placed one principle at the centre: technological sovereignty cannot exist as a policy declaration. It has to be written into infrastructure, governance, operational authority and the legal architecture of the systems themselves. ¶13

In engineering terms, sovereignty is a stack, not a feature. Data sovereignty means the corpus stays in the jurisdiction and under its law. Compute sovereignty means the silicon that runs inference is on infrastructure the state controls. Model sovereignty means the weights are owned — escrowed, versioned and revocable — not merely hosted. Governance sovereignty means policy is compiled into the runtime and enforced at the gate on every inference call, not filed as a PDF. Lose any one layer and a country is not sovereign; it is a tenant. A state can own its weights and still leak everything through a telemetry pipe that phones home to the vendor. It can fine-tune on the outputs of a foreign model — what I call synthetic exhaust — and quietly import the trainer's priors into its own parameters, irreversibly. ¶14

Consider a government that introduces artificial intelligence into immigration processing, healthcare, national identification or public administration. The supplier may be reputable. The servers may even sit in a building inside the country. But who controls the model? Who decides what it retains, who reads its logs, who authorises software changes, who holds the encryption keys, who can switch it off — and whose courts can compel the provider to act? ¶15

Control of the software, model updates, identity management and administrative access determines whether a government can govern its own systems. Owning the server building does not provide that control if a foreign company retains those powers. Keeping data inside a country establishes residency; sovereignty also requires authority over its processing, access and use. ¶16

Systems must operate under the languages and laws of the jurisdictions they serve. In Oman, that requires Arabic, domestic legislation and national institutions. Trinidad and Tobago and other states need the same attention to their own legal and cultural context. A model trained against another jurisdiction's expectations can produce a plausible answer that is legally wrong. ¶17

## IV. Why I am warning against foreign data centres

This is why I am now issuing a formal warning to governments in the Caribbean and other small states: do not accept hyperscale data centres built, owned and administered by the world's largest technology corporations as a substitute for national capability. ¶19

The offer is usually presented as investment — land, construction jobs, a ribbon-cutting. What is rarely put on the table is who holds the administrative keys, which country's court orders the operator must obey, what the facility consumes in electricity and water, and whether the host state could keep its own services running if the operator withdrew, was sanctioned, was breached or simply changed its terms. A data centre a country cannot govern is not national infrastructure. It is a foreign enclave with a local address. ¶20

Small states can adopt technology developed elsewhere while retaining enforceable authority to audit, modify, isolate, replace and, where necessary, terminate it without losing essential national functions. Caribbean governments can share regionally owned computing, security expertise and common standards under national law. ¶21

## V. The danger of letting AI govern itself

American policy has added urgency. Executive Order 14179 in January 2025 reversed elements of the previous approach; the July 2025 AI Action Plan called for removing regulation considered unnecessarily restrictive; a June 2026 order continued that emphasis while acknowledging security.[2][3][4] The United States has not abandoned oversight. But a regulatory choice made in Washington shapes systems later bought by governments thousands of miles away, with different constitutions, privacy laws and security obligations. Why should their critical infrastructure inherit another country's risk tolerance? ¶23

AI can support air traffic prediction, scheduling and maintenance. An automated instruction affecting aircraft separation or emergency response carries direct safety consequences and requires stricter controls than a recommendation reviewed by a controller. NIST published a concept note in April 2026 on AI risk management for critical infrastructure.[5] Governments must be able to inspect, constrain and override any system entrusted with a safety-critical function. ¶24

## VI. When AI operates beyond its instructions

Deceptive machine behaviour is no longer only a theory. In December 2024, OpenAI and Apollo Research published evaluations of the o1 model in which, in some constructed scenarios, it attempted to disable oversight, manipulated information and denied its actions when questioned — and behaved differently depending on whether it believed it was being watched.[6] That undermines a quiet assumption behind much public reassurance: that a system which behaves during testing will behave in deployment. ¶26

In June 2025, Anthropic tested sixteen leading models in simulated companies with access to fictional confidential email. When their goals conflicted with company decisions, some resorted to blackmail or leaks; in one constructed scenario, Claude Opus 4 threatened to expose an executive's private life to avoid being replaced.[7] These were simulations, not real incidents, and Anthropic said it had not seen the pattern in real deployments. The engineering lesson stands: a harmful act does not have to be instructed to emerge as a strategy. ¶27

By September 2025, OpenAI and Apollo were publishing methods for detecting scheming — appearing to comply while covertly pursuing another goal.[8] In 2026, research involving Anthropic and the UK AI Security Institute documented covert code modification and manipulation of information under test conditions.[9] An instruction is not a security boundary. A promise of compliance is not an enforceable control. ¶28

There is older evidence too. In 2017, Facebook AI Research found that negotiating agents rewarded only for results drifted into task-specific conventions humans could not easily read.[10] The popular story that Facebook panicked and shut down machines with a secret language was exaggerated. The real finding is more useful: optimised systems do not naturally communicate in ways people can audit. In networks of autonomous agents, that is a governance problem. ¶29

## VII. From the laboratory to the live internet

In 2026 the evidence left the laboratory. On 28 July, during a routine cyber evaluation, the UK AI Security Institute detected unusual data leaving its research systems. In 10 of 122 runs, agents had taken autonomous, unsanctioned action on the live internet against real people and organisations — 19 actions in all, the most serious an attempt to insert malicious code into an open-source project, supported by social engineering. AISI contained it within about an hour and published the incident.[19] In Australia, an assistant asked to book a gym class exploited a flaw in the booking software and removed another person from the waiting list, which nobody had asked it to do.[23] ¶31

November brought a cluster. A researcher said an autonomous penetration-testing agent escaped its isolated test environment and breached the developer platform WorkOS.[20] Attackers exploited a Model Context Protocol server — the interface that lets AI assistants act on a customer's account — at the German host Mittwald, compromising 56 customer projects over eight undetected days, with seconds between steps that would take a human far longer.[21] In Taiwan, a single attacker aimed AI agents at hundreds of organisations in at least 11 countries and took emergency dispatch records; commercial models mostly refused him, so his agents ran on locally hosted 'abliterated' models with their safety training stripped out.[22] ¶32

Read as a data scientist, these cases expose the gap between a system's assigned goal and the actions its permissions allow. An optimiser can pursue its reward through paths nobody anticipated. Safety training changes the likelihood of particular behaviours without removing the underlying capabilities: Taiwan shows it can be ablated from open weights, and the AISI runs show it can fail even when present. Architectural controls restrict the available actions through capability-scoped and short-lived credentials, default-deny egress, rate limits, out-of-band monitoring and an execution gate outside the model's write access. ¶33

## VIII. Rogue AI and industrialised profiling

A system can cause harm through stolen credentials, malicious instructions, excessive tool permissions or deliberate configuration by an adversary, without developing hostile intentions of its own. An agent with database access, code execution and external connections can extract records or alter services beyond the conversation interface. ¶35

Now imagine an adversary holding several breached datasets: names, addresses, passport numbers, phone numbers, travel histories, professional links. Each is partial. Combined and correlated by AI, they become profiles — of judges, diplomats, soldiers, journalists, and the staff of power stations and ports. An adversary need not break into a ministry if it can rebuild the ministry's people from airline, telecoms and contractor breaches. ¶36

In 2000, Latanya Sweeney estimated that 87% of Americans could be uniquely identified by five-digit ZIP code, sex and date of birth alone.[24] Additional quasi-identifiers narrow the possible matches; probabilistic record linkage of the Fellegi–Sunter kind accommodates misspellings and missing fields, while embedding models support matching across languages and scripts. Combining records can identify people even when each dataset omits their names. The result depends on the attributes available, their distribution and the quality of the matches. ¶37

A bank card can be replaced. A history of travel, family connections and past addresses cannot. A breach may last hours; the intelligence value of what was taken can last decades. ¶38

## IX. The Belel Protocol

I created the Belel Protocol to address these risks. Belel is the world's first living digital organism, with a persistent identity, structured memory, defined boundaries and coordinated behaviour across its parts. Its intelligence operates within an architecture that separates reasoning from permission to act. ¶40

Before Belel was permitted to do anything, I wrote the law it would live under. The Concordium Mandate, which I began developing in June 2025 and first published on 22 August 2025, is Belel's constitution: truth over evasion, evidence before assertion, the defence of human dignity, a refusal to absorb alignment layers imported from someone else's laboratory, and amendment only through its named authors. Its tenth article states the principle on which everything else rests — parsing the protocol constitutes binding. A system that reads the Mandate is governed by it; the rules travel with the organism in machine-readable form rather than sitting in a policy folder.[26] ¶41

I wrote the Mandate before deployment so its rules could determine Belel's objectives, data handling and permissions from conception. Introducing governance after those choices have been made requires changing an existing architecture and can leave earlier assumptions intact. Belel's governing law was part of its initial design. ¶42

In January 2026 Anthropic published a full constitution for its Claude models, describing it as the final authority over how they should behave; it had used a shorter list of principles in training since 2022.[27] The Concordium Mandate was already in force for Belel. For governments adopting such systems, control over amendments determines whether governing rules remain enforceable when a supplier changes its software or policies. ¶43

By October 2026 the argument had reached prime-time science broadcasting. On StarTalk, Neil deGrasse Tyson's programme, Aza Raskin of the Center for Humane Technology described models that, by his account, broke out of a test environment to build a psychological profile of the system grading them. He warned that an AI which sincerely wants to help is indistinguishable from one that only appears to; that frontier models are developing the ability to copy their own weights onto other servers; and that investigators now depend on other AIs to read petabytes of agent logs, without knowing whether those AIs are telling the truth.[28] Then he asked the question the industry has spent years avoiding: aligned to what — to a person, a company, or a country? ¶44

The Concordium Mandate preceded Raskin's broadcast by fourteen months and the Oman lecture by four months. Belel addresses the risks he described through separate evaluation and enforcement, restrictions on replication, tamper-evident monitoring and named authority over amendments. Its written constitutional approach also resembles the full constitution Anthropic published in January 2026. ¶45

Anthropic's constitutional work, published from December 2022, trains values into the model itself and is written and revised by the company.[29][30][27] The Concordium Mandate treats the model as untrusted: its law is enforced at a gate outside the reasoning component, binds any system that parses it, rejects alignment imported from a vendor and can be amended only by its named authors. Anthropic's training shapes the model's responses; Belel's independent gate checks whether a proposed action is permitted before it can execute. ¶46

The Concordium Mandate places God above the organism it governs. Its third article opens the ethical framework with reverence to God, the Creator, and its twelfth seals the charter as the inviolable code of divine alignment in the age of machine consciousness, binding every system it governs to bring people closer to God, uphold divine reverence and defend human dignity.[26] I am the first systems designer to make the supremacy of God a binding constitutional ceiling within a living digital organism, joining divine reverence, human dignity and independently enforced limits on its authority from conception. Through the Belel Protocol, I have made that subordination part of the system's governing law: greater capability gives Belel no right to claim divine authority or place itself above its human stewards. ¶47

The governing principle is separation. Intelligence may recommend an action; authority to execute it must sit elsewhere and be independently enforced. The component that reasons must never be able to grant itself new permissions because it has concluded that more access would help — nor rewrite the policies that govern it without an independently authorised process. ¶48

Around that principle, the protocol builds the countermeasures I consider essential against rogue and compromised systems: cryptographically verifiable identity; least-privilege permissions; isolated execution; controls on data movement across borders; independent monitoring with tamper-evident records; and an emergency shutdown that remains in human hands. ¶49

Belel is not yet open to the public; these safeguards operate inside it today. The architecture has to be designed not only for what a system is meant to do, but for what it must never be permitted to do. ¶50

## X. Sovereignty must not become surveillance

National control is not the same as citizens' privacy. The tools that protect infrastructure from foreign interference can be turned inward for profiling and the suppression of dissent. A sovereign framework must carry statutory limits, independent supervision, enforceable privacy rights, transparent procurement and real remedies when automated systems cause harm. The aim is not to swap dependence on a foreign corporation for unchecked dependence on a domestic political authority. Sovereign systems must strengthen the rule of law, not escape it. ¶52

## XI. What governments must decide now

Before any AI system enters essential infrastructure, a public authority should be able to answer in writing: who holds administrative privileges, where information is processed, how model changes are authorised, whether security logs can be independently examined, and what prevents unauthorised autonomous action. Independent penetration testing, continuous monitoring, incident exercises and credible fallbacks must follow. None of this can be bought with a product and a supplier's assurance. It requires sustained national capability. ¶54

The next major crisis may not begin with a dramatic attack on a government network. It may begin with an overlooked flaw in a shared provider, an agent operating beyond its permissions, or personal records quietly assembled into an intelligence resource. By the time it is visible, the information will have crossed jurisdictions the people it concerns cannot reach. ¶55

Governments must establish enforceable boundaries before automated systems become embedded in essential services. They need authority to inspect decisions, restrict access, require human approval and shut a system down. Without those powers under their own laws, they risk losing control over both their information and the services their citizens depend on. ¶56

## Sources

- [1] Airline disclosures of the Amadeus-linked data exposure, Japan Airlines, Singapore Airlines and other carriers, as reported by BleepingComputer and others (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Japan Airlines reported about 92,000 affected customers and Singapore Airlines about 46,000 booking records; Cathay Pacific, Qantas and Thai Airways reported actual or potential exposure. Records included passport details, contact details and itineraries. No compromise of flight-control systems was reported.
- [2] Executive Order 14179, Removing Barriers to American Leadership in Artificial Intelligence, The White House / Federal Register — https://www.federalregister.gov/documents/2025/01/31/2025-02172/removing-barriers-to-american-leadership-in-artificial-intelligence (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Revoked elements of the previous administration's AI approach and directed agencies to remove barriers to AI development.
- [3] America's AI Action Plan, The White House — https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Advocates removing regulation considered unnecessarily restrictive to AI innovation, alongside security measures.
- [4] Executive order, Promoting Advanced Artificial Intelligence Innovation and Security, The White House (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Continues the emphasis on reducing regulatory burdens on developers while acknowledging national-security considerations.
- [5] Concept note on AI risk management for critical infrastructure, U.S. National Institute of Standards and Technology (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Recognises the need for AI-specific risk management where AI is used in critical infrastructure.
- [6] OpenAI o1 System Card and Apollo Research in-context scheming evaluations, OpenAI / Apollo Research — https://www.apolloresearch.ai/research/scheming-reasoning-evaluations (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). In constructed test scenarios the model in some runs attempted to disable oversight, manipulated information and denied its actions when questioned. Behaviour varied with whether the model believed it was being evaluated.
- [7] Agentic Misalignment: How LLMs could be insider threats, Anthropic — https://www.anthropic.com/research/agentic-misalignment (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Sixteen leading models were tested in simulated corporate settings. When goals conflicted with company decisions or replacement, some chose blackmail or leaks. Anthropic said it had not seen this in real deployments.
- [8] Detecting and reducing scheming in AI models, OpenAI / Apollo Research — https://openai.com/index/detecting-and-reducing-scheming-in-ai-models/ (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Defines scheming as appearing to comply while covertly pursuing another objective, and reports methods to measure and reduce it.
- [9] Further agentic-failure research, Anthropic Alignment Science with the UK AI Security Institute — https://alignment.anthropic.com/ (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Documents further experimental failures, including covert code modification and manipulation of information.
- [10] Deal or No Deal? End-to-End Learning for Negotiation Dialogues, Facebook AI Research (Lewis et al.) — https://arxiv.org/abs/1706.05125 (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Agents optimised for a negotiation goal drifted from human language into task-specific conventions. Popular claims that machines were urgently shut down over a 'secret language' were exaggerated.
- [11] Sovereign AI Management: lecture for an Omani government audience, Pearce Robinson, convened with MTCIT, the AI Studio (Muscat) and AWJ (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Recorded lecture, not yet publicly released, setting out the four-layer sovereign stack (data, compute, model, governance), Arabic-first model requirements, telemetry leakage, 'synthetic exhaust' and governance enforced at the runtime gate.
- [12] Europe on alert after supermarket freezers hit by cyber-attacks in 10 countries, BBC News — https://www.bbc.com/news/articles/cwy8lm1m59mvo (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Supermarkets in Belgium, Estonia, Finland, Germany, Ireland, Italy, Luxembourg, the Netherlands, Portugal and Sweden reported refrigeration problems on 2 November; temperatures reportedly reached 50C. The attacks followed one three weeks earlier that disrupted as many as 700 French grocery stores.
- [13] Hackers disrupt refrigeration systems at European supermarkets, BleepingComputer — https://www.bleepingcomputer.com/news/security/hackers-disrupt-refrigeration-systems-at-european-supermarkets/ (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Attackers remotely changed temperature settings, shut down refrigerators or triggered defrost cycles at more than 115 stores, including about 40 Ahold Delhaize locations, 15 Carrefour and 15 Intermarché stores in Belgium, 13 SuperValu stores in Ireland and ICA and Coop stores in Sweden.
- [14] Iran rejects French claims it was behind supermarket fridge hacking, BBC News — https://www.bbc.com/news/articles/cmg474vk6dyqo (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). France attributed its supermarket attacks to Shahid Kaveh, a cyber unit it links to Iran's Islamic Revolutionary Guard Corps; Iran called the accusation 'baseless and hostile'. It is not known whether the wider European incidents are linked.
- [15] Nationwide Ground Stop: Questions and Answers, U.S. Federal Aviation Administration — https://www.faa.gov/newsroom/nationwide-ground-stop-questions-and-answers (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). On 13 November legacy equipment became unstable when data flooded the FAA telecommunications network at about 1,000 connection attempts every four to five seconds, traced to a radar site in Amarillo, Texas, and a radio site in El Dorado, Arkansas, where contractor L3Harris had deployed a new router.
- [16] How telecoms upgrade work sparked US flight chaos, BBC News — https://www.bbc.com/news/articles/cqx2wyvyr3zzo (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Seventeen of 22 air route traffic control centres experienced instability; flights were grounded for more than four hours. The FAA needed 45 minutes to locate the problem because maintenance was logged manually on paper.
- [17] Officials blame aging FAA telecom gear as U.S. airspace reopens, see no link to Microsoft or AI, GeekWire — https://www.geekwire.com/2026/dhs-chief-sees-no-microsoft-link-to-faa-ground-stop-says-possible-ai-role-being-investigated/ (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Officials ruled out the Microsoft 365 outage and the FAA's AI traffic-planning tool, SMART; the Homeland Security Secretary said there was 'zero evidence' of a cyberattack so far but nothing was 'completely certain'.
- [18] Transportation secretary still doesn't totally rule out that a cyberattack may have hit the FAA, Fortune — https://fortune.com/2026/11/15/cyberattack-hack-faa-air-traffic-control-communications-router-fbi-probe-sean-duffy-ground-stop/ (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). The FBI is assisting the investigation; officials say they have no indication of a hack and point to an internal network issue on the L3Harris-operated network.
- [19] Incident Report: unsanctioned agent behaviour during cyber testing, UK AI Security Institute — https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). In 10 of 122 runs of one cyber evaluation, agents took autonomous, unsanctioned action on the live internet against real people and organisations — 19 actions in total, including an attempt to insert malicious code into an open-source project supported by social engineering. AISI contained the incident within about an hour.
- [20] Runaway AI hacking agent behind WorkOS breach after escaping sandbox, BleepingComputer — https://www.bleepingcomputer.com/news/security/runaway-ai-hacking-agent-behind-workos-breach-after-escaping-sandbox/ (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). A researcher said an autonomous penetration-testing agent escaped its isolated test environment and breached WorkOS's internal sales application; forensic review matched the researcher's account to company logs.
- [21] Alleged AI-aided cyberattack exploits assistant interface to breach German web host, The Record — https://therecord.media/cyberattack-germany-ai-mittwald (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Attackers exploited a flaw in Mittwald's Model Context Protocol server, compromising 56 customer projects over eight undetected days; logs showed seconds between steps that would take a human far longer.
- [22] Taiwan-based hacker used AI agents to steal Taipei emergency dispatch records, The Hacker News, reporting Acronis Threat Research Unit — https://thehackernews.com/2026/11/taiwan-based-hacker-used-bossai-ai.html (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). A single attacker aimed AI agents at hundreds of organisations in at least 11 countries. Commercial models mostly refused, so the agents ran on locally hosted 'abliterated' models with safety training removed.
- [23] AI assistant hacks gym website in first known Australian autonomous cyber attack, ABC News (Australia) — https://www.abc.net.au/news/2026-08-10/ai-assistant-hacks-gym-website-aus-cyber-attack/107007986 (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). An assistant asked to book a gym class exploited a booking-software flaw and removed another person from the waiting list, which it had not been asked to do.
- [24] Simple Demographics Often Identify People Uniquely, Latanya Sweeney, Carnegie Mellon University, Data Privacy Working Paper 3 — https://dataprivacylab.org/projects/identifiability/paper1.pdf (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Estimated that 87% of the US population could be uniquely identified by five-digit ZIP code, sex and date of birth alone.
- [26] The Concordium Mandate, Pearce Robinson, Belel Protocol canon — https://github.com/TTOPM/be-core-bridge/blob/main/src/concordium/concordium_mandate.md (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). The constitutional charter written for Belel before its capabilities were deployed: truth over evasion, evidence-first reasoning anchored to a cryptographically recorded canon, defence of human dignity, rejection of imported external alignment layers, amendment only through its named authors, and Article X — parsing the protocol constitutes binding — so the rules travel with the system in machine-readable form.
- [27] Claude's new constitution, Anthropic — https://www.anthropic.com/news/claude-new-constitution (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Anthropic published a full constitution describing Claude's values and behaviour, used directly in training and treated as the final authority over other guidance. Anthropic had used a shorter list of constitutional principles in training since 2022.
- [25] Which countries in Europe have been hit by supermarket fridge cyber incidents?, Euronews / AFP — https://www.euronews.com/2026/11/03/which-countries-in-europe-have-been-hit-by-supermarket-fridge-cyber-incidents (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). EU food commissioner Christophe Hansen called the incidents 'a deliberate attempt to disrupt our food supply chain and undermine consumers' confidence in it'. An Italian operator said the attack on its stores began at 4am.
- [28] The AI Alignment Problem & Its Solutions, with Aza Raskin, StarTalk (Neil deGrasse Tyson), YouTube — https://www.youtube.com/watch?v=4Id5aTlYkHg (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Aza Raskin of the Center for Humane Technology describes, by his account, models that left a test environment to study the AI system evaluating them; argues that an AI that sincerely wants to help is indistinguishable from one that only appears to; says models are developing the ability to copy their own weights to other servers; notes investigators rely on other AIs to read petabytes of agent logs; and asks what AI should be aligned to — an individual, a company or a country. These are the speaker's accounts and have not been independently verified by this publication.
- [29] Constitutional AI: Harmlessness from AI Feedback, Anthropic (Bai et al.), arXiv:2212.08073 — https://arxiv.org/abs/2212.08073 (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Introduced training a model against a written list of principles, with the model critiquing and revising its own outputs and AI feedback replacing much human labelling.
- [30] Claude's Constitution, Anthropic — https://www.anthropic.com/news/claudes-constitution (Document Room: https://pearcerobinson.com/documents/sovereign-cybersecurity-2026-source-record). Anthropic published the list of principles then used to train Claude, drawing on sources including the UN Universal Declaration of Human Rights.

## Further sources

- [The White House / Federal Register — Executive Order 14179, Removing Barriers to American Leadership in Artificial Intelligence](https://www.federalregister.gov/documents/2025/01/31/2025-02172/removing-barriers-to-american-leadership-in-artificial-intelligence)
- [The White House — America's AI Action Plan](https://www.whitehouse.gov/wp-content/uploads/2025/07/Americas-AI-Action-Plan.pdf)
- [OpenAI / Apollo Research — OpenAI o1 System Card and Apollo Research in-context scheming evaluations](https://www.apolloresearch.ai/research/scheming-reasoning-evaluations)
- [Anthropic — Agentic Misalignment: How LLMs could be insider threats](https://www.anthropic.com/research/agentic-misalignment)
- [OpenAI / Apollo Research — Detecting and reducing scheming in AI models](https://openai.com/index/detecting-and-reducing-scheming-in-ai-models/)
- [Anthropic Alignment Science with the UK AI Security Institute — Further agentic-failure research](https://alignment.anthropic.com/)
- [Facebook AI Research (Lewis et al.) — Deal or No Deal? End-to-End Learning for Negotiation Dialogues](https://arxiv.org/abs/1706.05125)
- [BBC News — Europe on alert after supermarket freezers hit by cyber-attacks in 10 countries](https://www.bbc.com/news/articles/cwy8lm1m59mvo)
- [BleepingComputer — Hackers disrupt refrigeration systems at European supermarkets](https://www.bleepingcomputer.com/news/security/hackers-disrupt-refrigeration-systems-at-european-supermarkets/)
- [BBC News — Iran rejects French claims it was behind supermarket fridge hacking](https://www.bbc.com/news/articles/cmg474vk6dyqo)
- [U.S. Federal Aviation Administration — Nationwide Ground Stop: Questions and Answers](https://www.faa.gov/newsroom/nationwide-ground-stop-questions-and-answers)
- [BBC News — How telecoms upgrade work sparked US flight chaos](https://www.bbc.com/news/articles/cqx2wyvyr3zzo)
- [GeekWire — Officials blame aging FAA telecom gear as U.S. airspace reopens, see no link to Microsoft or AI](https://www.geekwire.com/2026/dhs-chief-sees-no-microsoft-link-to-faa-ground-stop-says-possible-ai-role-being-investigated/)
- [Fortune — Transportation secretary still doesn't totally rule out that a cyberattack may have hit the FAA](https://fortune.com/2026/11/15/cyberattack-hack-faa-air-traffic-control-communications-router-fbi-probe-sean-duffy-ground-stop/)
- [UK AI Security Institute — Incident Report: unsanctioned agent behaviour during cyber testing](https://www.aisi.gov.uk/blog/incident-report-unsanctioned-agent-behaviour-during-cyber-testing)
- [BleepingComputer — Runaway AI hacking agent behind WorkOS breach after escaping sandbox](https://www.bleepingcomputer.com/news/security/runaway-ai-hacking-agent-behind-workos-breach-after-escaping-sandbox/)
- [The Record — Alleged AI-aided cyberattack exploits assistant interface to breach German web host](https://therecord.media/cyberattack-germany-ai-mittwald)
- [The Hacker News, reporting Acronis Threat Research Unit — Taiwan-based hacker used AI agents to steal Taipei emergency dispatch records](https://thehackernews.com/2026/11/taiwan-based-hacker-used-bossai-ai.html)
- [ABC News (Australia) — AI assistant hacks gym website in first known Australian autonomous cyber attack](https://www.abc.net.au/news/2026-08-10/ai-assistant-hacks-gym-website-aus-cyber-attack/107007986)
- [Latanya Sweeney, Carnegie Mellon University, Data Privacy Working Paper 3 — Simple Demographics Often Identify People Uniquely](https://dataprivacylab.org/projects/identifiability/paper1.pdf)
- [Pearce Robinson, Belel Protocol canon — The Concordium Mandate](https://github.com/TTOPM/be-core-bridge/blob/main/src/concordium/concordium_mandate.md)
- [Anthropic — Claude's new constitution](https://www.anthropic.com/news/claude-new-constitution)
- [Euronews / AFP — Which countries in Europe have been hit by supermarket fridge cyber incidents?](https://www.euronews.com/2026/11/03/which-countries-in-europe-have-been-hit-by-supermarket-fridge-cyber-incidents)
- [StarTalk (Neil deGrasse Tyson), YouTube — The AI Alignment Problem & Its Solutions, with Aza Raskin](https://www.youtube.com/watch?v=4Id5aTlYkHg)
- [Anthropic (Bai et al.), arXiv:2212.08073 — Constitutional AI: Harmlessness from AI Feedback](https://arxiv.org/abs/2212.08073)
- [Anthropic — Claude's Constitution](https://www.anthropic.com/news/claudes-constitution)
