top of page
Glossy 3D 48PEGASUS Logo.png

Cyber Risk from Early Hacking to Agentic AI

Incident casebook and recovery

[*REDACTED*] Worm: the limits of an informal Internet

1988 · Availability shock · Federal conviction

Known event and detection

On November 2, 1988, a self-propagating program overloaded computers on the early Internet. The FBI describes an estimate of roughly 6,000 affected machines, but the historical figures are not a complete host census or an appropriate denominator for today’s networks. Administrators saw critical computing resources consumed and services become unusable. The harm came from compounding behavior, not a requirement for each victim to be individually chosen. [41]

What helped stop disruption

Administrators interrupted network connectivity and collaborated on countermeasures in a setting with fewer formal reporting channels than today. Separately, the FBI conducted an investigation after the incident became public. [*REDACTED*] was convicted under the Computer Fraud and Abuse Act; the investigative record should not be treated as proof of every technical account circulated after the event. [41]

What rebounded

The incident helped drive formal coordination across institutions. The FBI describes the creation of the first computer emergency response team after the event. The lasting change was not the removal of one program but the expectation that organizations should share incident information and have a prepared response contact. Individual restoration times are not consistently recorded in the public case history. [41]

Present relevance

The equivalent design risk today is any automated process that can act on many systems without independent rate, scope, and stop controls. Contain effects before establishing perfect attribution. Self-propagation is a historical fact here, not an analogy establishing that contemporary AI already behaves in the same way. Exercises should verify that a failing service can be paused without relying on the same disrupted communications channel.

Mail-era worms: why human trust became a system risk

Early mass adoption · Social trust · Coordination

What changed

The spread of electronic mail made familiar interpersonal signals part of computing risk. In May 2000 the “Love Bug” caused worldwide disruption; the FBI’s congressional testimony later described it in the history of Internet worms. Modern communications carry more information and reach more devices, but the useful comparison is misplaced trust in routine material, not any particular historical technical mechanism. [42]

Recognition under noisy conditions

A sharp rise in similar reports and abnormal service demand can point to a common problem, even when users experience it differently. A responder should preserve a few well-scoped observations before declaring that every anomaly has the same cause. Early official retrospective sources do not provide comparative telemetry for all affected organizations, so they cannot establish typical detection or recovery time. [42]

Containment and recovery

Defenders used coordination and protective controls appropriate to the affected communication systems; once isolated disruption was understood, service restoration had to avoid resuming the same harmful behavior. For current organizations, an approved alternate notification channel lets teams reach users even when normal messaging is impaired. A controlled return to service also helps separate residual queued work from new warning signs. [42] [78]

Modern parallel

Synthetic voice and text can make deception more persuasive, but older incidents already showed that a familiar identity is not the same as authority. Verify consequential requests in a trusted channel and make clear who may approve account, payment or access changes. Do not make a user’s ability to recognize unfamiliar phrasing the sole safety control.

Exhibit 03 · From network incidents to shared-service failures

Selected cases illustrate shifts in the services and dependencies affected. Milestones mark incident years, not a complete history or consistent measure of harm. Placement is in chronological order rather than scaled to elapsed years. See the historical casebook and its original sources.

48PEGASUS Research timeline; individual primary records are cited in the casebook.

[*REDACTED*]: payment security and an outside signal

2013 · Retail payments · Distinct affected categories

What the witness said

[*REDACTED*] told Congress that the Justice Department notified it of suspicious payment-card activity on December 12, 2013. An outside signal can identify consequences that an organization cannot see from its own application logs. The witness identified approximately 40 million customers affected in the payment-card category and up to 70 million in the separate personal-data category. The populations may overlap; they should not be added as distinct people. [43]

Why scope mattered

The business had to determine what happened, which channels and periods were affected, and what protection customers and financial institutions needed. A finding about retail payment records does not automatically describe online transactions or establish how every individual was harmed. Broad references to third-party access belong in a control discussion, not an incident recreation.

From containment to remedy

Target testified that it mobilized outside expertise and acted to stop the unauthorized activity, working with law enforcement and financial partners. Follow-up and customer communications continued after immediate containment. Those are separate milestones: compromised activity stopped, payment services stable, and customer remedial actions underway do not mean the same thing. The earlier witness account should not be mistaken for a final loss estimate. [43]

Modern parallel

Give third-party access a defined purpose and expiration; name the organization that reviews unusual activity. If a bank, partner or regulator calls with a plausible warning, on-call leaders should be able to verify the caller and rapidly compare incident observables. Third-party access should be measured against the data and operations reachable, not just against whether the partner has a contract. [78]

OPM: an incident may outlast the exposed system

2015 · Federal records · Privacy recovery

Discoveries and scope

The Office of Personnel Management identified a personnel-data breach in April 2015 involving about 4.2 million current and former federal employees. It then reported a separate, related incident in background-investigation systems involving 21.5 million people. The second group included some people outside the federal workforce. Neither figure is a measure of how many people subsequently experienced fraud; combining the two without de-duplication is misleading. [44] [45]

Response and uncertainty

OPM notified affected people and described an interagency forensic response. Its contemporaneous notice assessed that the adversary was no longer active in its network based on the evidence available at that time. The conclusion did not imply that previously copied data were returned, no privacy risk remained, or other affected environments were independently certified clean. [44]

Rebound measured against verified controls

GAO’s 2017 follow-up found 11 of 19 US-CERT recommendations completed; action had been taken on the other eight, with four of those actions requiring further improvement. GAO also found weaknesses in tracking expected completion dates and validating effectiveness. A later 2024 review examined different systems under DOD’s Defense Counterintelligence and Security Agency and identified continuing governance and privacy-control needs. This was not a finding that the original incident continued. [45] [46]

Modern parallel

Old records can remain sensitive long after a service is rebuilt. Minimize retention where authorized, impose stronger controls on exceptional access, and define how downstream partners must respond when identity data have been exposed. Report corrective-action progress with evidence that a safeguard works, not a count of closed tickets.

Mirai: devices affected elsewhere can interrupt your service

2016 · DDoS · Shared internet dependency

What happened and how it was recognized

In 2016, networks of compromised connected devices were involved in denial-of-service incidents that affected downstream internet services. Justice Department accounts distinguish later individual pleas and investigations, including a 2020 plea relating to attacks against Dyn and other targets, from the earlier cases about originating botnets. A service provider can experience interruption even when most participating devices and the systems that rely on its service are operated by different parties. [47] [48]

Detection is local; the failure is distributed

Operators and users encounter slow or unavailable service, while providers can distinguish excess traffic and breakdowns of expected demand patterns. These official criminal case records establish roles and consequences, not complete timing at every downstream company. They should not be cited as though all outages had one source or every participating device was known.

What rebounded

Providers mitigated the disruptions and justice authorities arrested or prosecuted people in distinct investigations. Restoration of a provider’s service is separate from cleanup of the broad connected-device population. A guilty plea establishes the specific conduct admitted, not responsibility by all the parties covered across multiple proceedings. The broader dependency persisted, creating incentives for better device lifecycle management and continuity design. [47] [48]

Modern parallel

Inventory dependency on authoritative internet services and assess whether a supposed alternate has different upstream infrastructure. Exercise failover before it is needed and verify consumers can find the alternate. Device defaults and update support are governance issues as much as technical properties. A healthy application cannot overcome every failure at its dependencies.

WannaCry: restoration is a health-service outcome

2017 · England NHS · Outage versus clinical effects

Detection and immediate response

The UK National Audit Office describes the May 2017 WannaCry disruption as the first ransomware attack of this scale experienced by the NHS. Multiple services independently recognized disruption and the health system coordinated a response. Organizations with unsupported or unpatched software were susceptible; exposure did not mean that all such organizations were infected. Some organizations shut down services as a precaution, so received infection and operational disruption are different denominators. [49] [50]

Measure the disruption accurately

NAO estimated more than 19,000 appointments would have been cancelled based on partial data. Its review found that some emergency services diverted patients, while emergency care continued. Unknown follow-up should not be presented as proof of no patient harm. The later NHS review noted nuances and challenges classifying organizational impact. [49] [50]

Recover care rather than devices alone

NHS services used continuity measures and restoration work, but the status of a hospital workstation was not a complete measure of care recovery. A patient still needs access, an appointment rebooked where necessary, and a reliable record of care delivered while systems were unavailable. For a continuity dashboard, pair device recovery with cancelled and rescheduled appointments, delayed results and affected locations. [49] [50]

Control changes

The reviews emphasize incident coordination, maintained systems, cyber alerts implemented and verified, and recovery exercises. Patch delivery is not the same as protective deployment, and patching is not a substitute for service continuity. Test fallback with the workers expected to operate it and plan for reconciliation when digital systems return.

NotPetya: destructive incidents can travel across borders

2017 · Destructive malware · Operational continuity

Scope and attribution

NotPetya caused broad disruption in June 2017. U.S. prosecutors later charged six Russian military intelligence officers in connection with destructive operations, including NotPetya. The charges are allegations, not a conviction of all defendants. [*REDACTED*]’s annual report independently confirms the company’s global business disruption and estimates approximately $250–300 million in company losses across such categories as lost revenue and restoration and operating expense. It is not an estimate of world totals. [51] [52]

Visibility under shared failure

When multiple business units lose service together, lack of an independent administrative or recovery channel can complicate investigation. The high-level record supports widespread destruction and substantial cross-border effects. It does not justify treating every company’s outage or response as identical. The recognized harm also did not depend on every affected entity being a specifically intended target.

What rebound means here

[*REDACTED*]’s annual report records both operational effects and restoration costs; a narrower service milestone should not be described as zero remaining cost. Defensive teams need a known-good restoration authority, clean copies, and independent ways to confirm business transactions. A demand for payment does not establish that complete recovery is possible through payment; this episode is an example of destructive behavior. [51] [52]

Modern parallel

Treat recovery architecture as a dependency graph: access to recovery materials, system configuration, and responsible personnel should not all disappear when one enterprise system fails. Maintain an offline resource and alternate approval process. Exercise recovery in priority order rather than measuring success by the number of servers restarted.

[*REDACTED*]: verified remediation beats a patch bulletin

2017 · Personal data · GAO and FTC records

Control failures and detection

GAO identified shortcomings in identifying and correcting a vulnerable system, enforcing segmentation, and detecting or limiting access to sensitive data. [*REDACTED*] became aware of suspicious behavior in late July and began its response. This case should be read as multiple safeguards failing together, not solely a missing patch or a judgment about any one staff member. [53]

Different counts at different dates

GAO reported at least 145.5 million people whose personal information was accessed based on the evidence it reviewed. The FTC’s later settlement announcement refers to approximately 147 million. The difference is not an arithmetic error or two separate breaches at the company; each publication used information available within its own scope and time. [53] [54]

Restoring trust and addressing harm

The company acted to investigate, contain and remediate; authorities separately pursued remediation obligations. Regulators announced a settlement of at least $575 million and potentially up to $700 million. Those amounts are settlement terms, not a measure of all societal costs, a maximum imposed fine alone, or an amount proven disbursed to all victims. The oversight record distinguishes preventive obligations from consumer relief. [53] [54]

Lesson for today

Track critical exposure against an inventory you can verify. Assign each exception a time limit, business owner and alternate mitigation. Confirm visibility is working end to end, including expiration management and an independent check that expected monitoring data still arrive. After an incident, measure time to customer protection as well as time to secure the service.

[*REDACTED*]: when the trusted supplier is the multiplier

2020 · Supply chain · Exposure is not intrusion

Detection and incident scale

A defender’s investigation pointed to compromise linked to a trusted enterprise software product. GAO and congressional testimony distinguish the estimated 18,000 customers to whom compromised software was sent from the smaller group assessed as follow-on victims. The FBI testified in 2021 that it had then identified nine federal agencies and fewer than 100 nongovernment entities in that latter category. Dates and investigation scopes matter; these are not final, universally comparable counts. [55] [56] [57]

Response under uncertainty

Authorities established a Cyber Unified Coordination Group, and CISA directed federal civilian agencies to disconnect or power down affected products. An immediate safety directive limited exposure but did not establish that all downstream systems had been analyzed. A software version list and a full incident impact assessment answer different questions. [55] [57]

Recovery and post-incident change

Restore confidence in software, identities and connected information, not only the compromised product. GAO highlighted communication, information-sharing and evidence challenges and reported incomplete implementation of foundational supply chain practices. Remediation of a customer endpoint cannot substitute for an independent evaluation of supplier controls or the integrity of downstream systems. [55] [56]

AI ecosystem parallel

An internally approved model, integration, or update can carry substantial delegated authority. Stage changes, maintain independent monitoring, require provenance, and keep emergency access independent of the component being assessed. Even a trusted signature should not replace evidence that the supplier and approved artifact are operating as expected. This is a governance analogy, not a claim of an equivalent AI attack.

[*REDACTED*]: operating safely under uncertainty

2021 · Critical infrastructure · Restart is not seizure

What was detected

[*REDACTED*] reported a ransomware incident in May 2021. Its CEO testified to rapidly stopping pipeline operations to contain the incident and assess the threat. A halt in physical operations does not by itself establish compromise of pipeline controls. In this case, the decision to stop was part of the safety and containment response, and the public record distinguishes the business disruption from confirmed intrusion scope. [58]

What the restart meant

Pipeline restart began on May 12 according to [*REDACTED*]’s testimony. Restarting a network of infrastructure is not instantaneous service normalization; it requires operations to verify milestones and communicate which constraints remain. Care in public wording matters when downstream supply and customer behavior may respond to incomplete information. [58]

Separate criminal recovery

DOJ later announced seizure of 63.7 bitcoin then valued at approximately $2.3 million and alleged they represented proceeds of the ransom payment. This is a case-specific enforcement result, not a guarantee of reclaiming payments or a measure of pipeline recovery, business losses, or value on the payment date. Early evidence preservation and reporting serve a different purpose from service restoration. [59]

Modern parallel

A halt decision should be rehearsed with people responsible for safe operation. Define confidence needed for restart, who signs it off, and how billing, dispatch, stakeholder messages and physical safety are verified. Ensure an emergency communication method is not solely dependent on the affected enterprise network. A post-event milestone should state clearly what is back and what remains uncertain.

Ireland’s HSE: continuity across an entire care network

2021 · Conti · Independent review

From disruption to resilience problem

Ireland’s Health Service Executive experienced a major cyber incident on May 14, 2021. The commissioned independent review details how local signs, incident escalation, and national coordination interacted. It identified shortcomings in cybersecurity maturity and ownership, and its evidence demonstrates the importance of acting on warnings even when nobody initially has a complete national view. Not every service or information system was affected in the same way. [60]

Containment costs can be necessary

Defenders acted to reduce exposure across the environment, interrupting service availability while the impact was assessed. For providers, computers and connectivity are not peripheral: diagnostics, referral coordination and appointment flows can all depend on shared systems. The patient-facing consequences must be evaluated independently of the evidence collected for technical analysis. [60]

Rebound and verification

Technical recovery required a coordinated restoration process across critical services; organizational remediation continued well after the initial emergency. HSE governance records show subsequent monitoring and remediation discussions. An agenda discussion documents governance attention, not completion of a control or independently measured reduction in risk. [60] [61]

Present relevance

Give local clinicians fallback pathways and a central process for setting recovery priorities. Record time to safe manual service, backlog and reconciling records, and authority to resume digital workflows. When evidence is incomplete, tell staff what they are permitted to do and what must wait. Rehearse a multi-site outage with external partners rather than relying on a technical tabletop alone.

Storm-0558: the customer noticed what the provider missed

2023 · Cloud identity · Independent observations

What public review established

In May and June 2023 a threat actor known as Storm-0558 accessed Exchange Online mailboxes affecting more than 500 people across 22 organizations according to the Cyber Safety Review Board. The review emphasizes deficiencies in cloud identity safeguards, the stewardship of authentication trust, and provider detection. U.S. State Department personnel identified suspicious activity using available audit data and brought it to [*REDACTED*]’s attention. [62]

Important uncertainty

The board reported that [*REDACTED*] was unable to determine how the relevant trust credential originally came under attacker control. That gap should remain in the account. Knowing what permission boundary failed does not reconstruct the cause of its original compromise. Initial corporate explanations and later independent findings should be presented in their proper date and scope. [62]

Response and recovery

[*REDACTED*] mitigated the incident and worked with affected organizations, while the review documented security and reporting improvements that should follow. An organization needs to consider earlier activity and residual exposure even after a provider invalidates the credential or fixes the affected service. Protecting confidentiality cannot be reduced to restoring service availability. [62]

Modern parallel

Do not place all evidence and all authority inside the same cloud trust boundary. Confirm contractually what logs are available, which events they cover, how long you can access them and how external evidence is preserved. Periodically exercise coordinated revocation and recovery with vendors. The lesson is independent verification, not that a modern authentication mechanism always fails.

[*REDACTED*]: restoring software does not un-disclose data

2023 · File transfer · Vendor and tenant impacts

A product finding and many customer questions

[*REDACTED*] disclosed in its fiscal 2023 Form 10-K response to a [*REDACTED*] security incident affecting transfer products. The company stated it remained operational and had not uncovered evidence at the time of filing of related unauthorized activity in its corporate environment. That statement is not reassurance that customers were unaffected. A service provider, a large customer, and people whose information passed through the service can have materially different incident scopes. [63]

Discovery, communication and response

The provider described discovering a problem, mobilizing its response, mitigating affected cloud deployments and supporting customers. Because customer environments varied, a patch notice alone could not answer whether a specific dataset had been exposed. An organization must reconstruct its own data movement and preserve the basis of any notice it sends. [63]

Rebound and legal uncertainty

[*REDACTED*]’s later filings continued to distinguish related costs, insurance and proceedings. The company announced that SEC staff had concluded its investigation without recommending enforcement; that statement about one investigation is not a finding that the incident produced no loss or that different proceedings were resolved. [64] [65]

Modern parallel

File transfer, cloud connectors and agent tools can expose material supplied by many separate organizations. Keep destination and retention inventories, prefer minimum necessary data, and know how to contact downstream owners. Response should reserve separate evidence lanes for product restored, exposure assessed, and notices or customer remedies completed.

[*REDACTED*]: payment outages become care risks

2024 · Claims clearinghouse · Impact extended into 2025

Recognition and immediate containment

[*REDACTED*]’s CEO testified that [*REDACTED*] isolated affected systems after discovering the February 2024 cyberattack. His testimony pointed to inadequate multifactor coverage on part of the affected environment. The outage interrupted services that pharmacies and clinicians relied on for claims and related transactions. This is an example of concentrated business dependence, not proof that every customer’s network was compromised. [66] [68]

Two impact denominators

AHA testimony described operational and cash-flow strain across providers. HHS states [*REDACTED*] notified its Office for Civil Rights on July 31, 2025 that approximately 192.7 million individuals had been impacted. The number is a reported affected-person estimate, not the number of provider systems compromised, denials of care, paid claims, or a measurement of subsequent identity misuse. [68] [67]

Restoration and the cash-flow tail

Public testimony documents staged restoration. Claims processing or eligibility checks coming back online do not mean delayed payments have reached providers or transactions have been reconciled. Business continuity requires testing alternative clearing routes, identifying the backlog that needs human resolution, and documenting whether smaller providers have adequate short-run liquidity. [66] [68]

Modern parallel

Measure single-vendor concentration in transactions, settlement timing and fallback capacity. Do not count the alternate as independent without checking its own underlying providers. Define service-level evidence and practical notice commitments with partners before an outage. Operational and privacy remedies can continue on entirely different clocks.

[*REDACTED*]: a counterexample to attack-by-headline

2024 · Nonmalicious outage · Concentration risk

Observed impact, distinct cause

[*REDACTED*] reported in a July 2024 SEC filing that a sensor configuration update caused outages affecting certain Windows systems and stated that the event was not caused by a cyberattack. [*REDACTED*] estimated roughly 8.5 million affected Windows devices. That estimate concerns devices, not distinct people, separate companies, or the number of services interrupted. The continuity lesson is valid even though adversarial intent was absent. [69] [71]

How a routine change becomes systemic

A vendor update can reach many customers in a short period. The official root-cause summary describes issues in validation and handling of the update, and remediation required more than normal software scheduling for affected systems. An incident commander should maintain alternative hypotheses until there is evidence distinguishing malicious activity from faulty deployment. [70]

Rebound and controls

The company reversed the update and developed further mitigations. Organizations still had to restore individual systems and essential workflows; a vendor’s rollback is not necessarily equivalent to endpoint recovery. Staged deployment rings, independent stop authority, out-of-band communications and rollback tests address the concentration risks common to malicious and accidental incidents. [69] [70]

Modern parallel

AI agents can also change many records rapidly under legitimate credentials. Broad reach should require rate limits, narrow scope and auditability independent of the agent. When the same provider controls deployment and telemetry, verify that an internal route can still stop the workflow and capture evidence if its monitoring fails.

Salt Typhoon: visibility in infrastructure shared by others

2024–2025 disclosures · Telecommunications · Campaign uncertainty

How to describe the public record

The FBI publicly described PRC-affiliated activity known as Salt Typhoon and a compromise of multiple U.S. telecommunications companies. Coordinated government statements and a multinational hardening publication emphasized visibility and protection. The presence of this official attribution is not permission to extend it to every telecommunications disruption, infer undisclosed victims or identify unsupported attacker names. [72] [73]

Detection is not the same as attribution

Network operators need useful telemetry before any specific actor is named. A provider can identify suspicious changes and constrain privileged actions without knowing which government, group or person is responsible. Full discovery chronology and customer-level victim scope are not consistently public, and the absence of such detail should not be filled with conjecture. [72]

Recovery and unfinished questions

Authorities published defensive guidance and sought additional investigative information. This supports response and mitigation but does not verify eradication across all providers, customer notification in all affected areas, or full rebuilding of trust. Individual provider controls require evidence that they are in place, reviewed and independently testable. [72] [73]

Modern parallel

For essential communications, maintain a vetted alternative channel and know which routes it ultimately uses. Define accountable owners for audit collection, review of administrative authority, and fallback restoration. Focus on how to discover unauthorized changes quickly without assuming normal service availability proves communications confidentiality.

Hive and LockBit: disruption does not erase risk

2023–2024 · Law enforcement and recovery · Separate measures

What was interrupted

DOJ reported a coordinated Hive disruption in January 2023, including supplying decryption assistance to victims. It estimated that more than $130 million in ransom demands were averted; this is an agency estimate of avoided payments, not the total cost of affected incidents or the amount of stolen data recovered. DOJ and UK authorities announced coordinated LockBit disruption in February 2024 and offered victim assistance. The two cases describe different operations and populations. [74] [75]

Why counting cases is not enough

Investigative success, service recovery and prevention are separate measures. Some victims may regain files while still incurring investigation and data-exposure costs. Counts of cases, demand amounts, infrastructure seizures and later convictions are not interchangeable. DOJ’s inspector general examined the difficulties of measuring program effectiveness, reinforcing the need to interpret individual announcements narrowly. [77]

Rebound on both sides

DOJ later brought allegations in the LockBit case and described attempts to revive the operation. Indictments allege conduct and should not be treated as verdicts. An interrupted group may not resume its former reach, while some people involved may continue in another operation. The appropriate result is bounded disruption, not a universal claim that the crime is finished. [76]

Present relevance

Offer employees and partners a straightforward path to report events, and prepare evidence-handling arrangements before systems are disrupted. A victim’s priority must still include independent restoration and data-scope determination. Following credible law-enforcement developments can reduce uncertainty, but no organization should predicate continuity on the availability of outside decryption help.

[*REDACTED*]: replenishment is not recovered property

October 29, 2026 · Financial controls · Dated and distinct estimates

Recognition and containment

[*REDACTED*] reported an internal alert about unusually large transfers and opened an investigation on October 29. The company said it suspended withdrawals, worked with partners to freeze assets, and rebuilt or replaced affected systems. This first-party account is provisional, and its description of completed mitigations is not a substitute for independent validation. An approved policy should identify who may suspend material transactions, what authority remains active, and how funds and authorizations are reconciled when service resumes. [79]

The issuer’s numbers answer different questions

[*REDACTED*]’s November 3 statement reports about $6.34 billion of assets impacted and roughly $2.85 billion frozen or intercepted, subject to reconciliation and changing valuations. It also stated that as of October 31 it had replenished 98% of affected token assets. On November 13 the company reported completing replenishment while continuing phased withdrawal resumption; some unsupported tokens were assigned a separate process. These are issuer statements. Replenishment can protect customers without returning stolen property, and nominal dollar valuations can move after the incident. [79] [81]

What the Justice Department established

On November 12 DOJ attributed the [*REDACTED*] theft to North Korean actors the FBI calls TraderTraitor, separately from another campaign described in the same release. It reported more than $2.7 billion frozen through collective efforts, including more than $1.2 billion within court-authorized seizure or restraint. The latter is nested within the former. Neither is a final forfeiture or return. Do not add platform and government figures; scope, timing, valuation, and overlap are not reconciled across the two accounts. [80]

Present relevance

Separate independent alerting from payment authority and use meaningful sign-off outside any automated agent proposing the transaction. Test whether stopping the front end also constrains actions already authorized elsewhere. Record separately which services are available again, which customers are made whole, and which property is restrained, returned, or still outstanding.

Exhibit 04 · Asset impact and legal restraint are not recovery

[*REDACTED*]’s November 3 estimate put impacted assets at about $6.34 billion. DOJ on November 12 reported more than $2.7 billion collectively frozen and over $1.2 billion under court-authorized seizure or restraint, nested within the frozen amount. Bars show the threshold for each DOJ figure, not the final balance. Customer replenishment and actual return of stolen assets are separate outcomes.

[*REDACTED*] initial incident update and DOJ November 12 release; see the casebook and source register.

How to tell whether an organization truly rebounded

Cross-case synthesis · Operational evidence · Verification gates

Distinguish four milestones

Containment prevents additional loss in the known scope. Technical recovery restores resources to a trusted state. Business recovery measures the service experienced by customers and staff. Corrective-action validation tests whether protective changes work and whether leadership owns remaining risks. A report can close one without closing the others. NIST frames incident response as part of ongoing cybersecurity risk management rather than a stand-alone emergency activity. [78]

What counts as evidence

For containment, document what remains isolated and which observations support that status. For technical recovery, use acceptance tests and independent monitoring. For business recovery, record throughput, remaining backlog, data reconciliation and impacted users. For validated remediation, name the independent tester, document scope and publish follow-up assignments and dates. OPM’s GAO follow-up illustrates why action completed is not the same as independently shown effective. [45]

Avoid the speed trap

Average restoration time compresses very different services and hides consequences for the last affected group. Report a timeline by function and denote any clock whose detection or completion date is unknown. Define whether the clock starts at first observed harm, formal declaration or discovery. If a safety or confidentiality concern remains unresolved, a fast restart may not indicate good performance. [49] [67]

Apply to agentic systems

A rollback must withdraw delegated authority and verify consequential actions already taken outside the tool. A model being offline does not reverse sent notices, altered records or disclosure of data. Require action logs, lower-risk mode and an independent way to pause execution. Historical cases show why evidence boundaries and business obligations matter even when the technology changes.

Exhibit 05 · From first signal to safe restoration

The incident lead coordinates containment and evidence preservation while service owners plan safe operation. Restoration and reopening are distinct decisions. Local safety, legal, operational and notification requirements can change the order.

48PEGASUS Research conceptual exhibit. Not empirical data.

What history can and cannot predict

Patterns, not frequencies · Unknowns are findings · Use counterexamples

Recurring lessons

The cases show repeated dependencies: unverified trust, inadequate isolation, missing independent observations, uneven recovery authority, and dilution of accountability across suppliers. They are selected for their explanatory value, not sampled to estimate current worldwide incidence. Cautions about concentrated controls remain valid even when the cause is not hostile, as the July 2024 [*REDACTED*] outage demonstrates. [55] [62] [69]

What cannot be inferred

The number of people affected, dollars estimated, devices disrupted, organizations assessed and funds seized are different evidentiary quantities. None can be combined into a single trend in malicious capability. Incident date, disclosure date and later report date also differ. Some cases have strong documentation on recovery and weak detail on discovery; others are the opposite.

Long-run planning

Use past incidents to test a contemporary decision: what happens if a provider sends a dangerous update, a trusted identity stops being reliable, or normal operations need an alternate? Identify the decision owner, evidence needed, expected recovery objective, and harm that restoration cannot reverse. Do not infer an inevitability or timeline for new forms of autonomy from past examples. [78]

Practical evidence discipline

When studying a current event, separate first-party reports, independent oversight findings, enforcement allegations and policy responses. If analyses conflict, name what each actually measures. Unknown initial access, how much data were copied, and when recovery ended are material facts, not gaps to fill with a representative story. This discipline makes case history useful without turning it into instructions for repeating an intrusion.

Previous chapter

Next chapter

Back to the report

Browse: White Papers · Reports · Downloads

bottom of page