top of page
Glossy 3D 48PEGASUS Logo.png

Cyber Risk from Early Hacking to Agentic AI

The contemporary threat landscape

Reported cyber losses are rising, but the ledger is partial

Complaints are not a census · Calendar year 2025 · Name the reporting channel

What the Bureau measured

The 2025 FBI Internet Crime Complaint Center annual report logged 1,008,597 complaints and $20.877 billion in reported losses, a reported 26% increase against the 2024 total. Rounded reference values for 2020 through 2025 are $4.2 billion, $6.9 billion, $10.3 billion, $12.5 billion, $16.6 billion and $20.877 billion. The reporting system records large financial losses and harmful conduct without requiring that they all be confirmed intrusions. [82] [83]

What the Bureau did not measure

This is the observed complaint channel, not the total cost of cybercrime or the number of unique victims. Awareness, willingness to report, labeling rules, referral practices, duplicate reports and changes in loss documentation affect the series. The presence of overseas reporting makes it unsuitable as a domestic population rate.

What the slope means

A rising complaint loss total establishes increasing reported harm in this record. It does not establish equal growth in technical sophistication or the fraction caused by a new technology. A rising count and a rising total can be driven by different groups or loss categories. An insurance series and a breach-response series should be charted separately because inclusion and accounting rules differ. [82] [85]

Operational implication

Use the national series to benchmark the scale of documented fraud, then estimate your organization’s own exposure with transaction patterns, protected populations, material assets and recovery tests. Register changes in reporting coverage before attributing a trend to adversaries. Show whether figures describe reported, confirmed, attempted, frozen, restored or paid amounts.

Exhibit 06 · Reported internet crime losses rose

The FBI received 1,008,597 IC3 complaints in 2025 involving $20.877 billion in reported losses, up 26% from 2024. Earlier values are rounded as presented by the FBI. Complaint-driven data do not capture all cyber-enabled harm and are not an attack incidence measure.

FBI, 2025 IC3 Annual Report: primary source

Financial deception dominates the recorded loss table

Same report, different categories · Trust is a control surface · Independent verification

The loss distribution

FBI reported approximately $8.649 billion for investment fraud, $3.047 billion for business email compromise, $2.135 billion for tech or customer support, and $1.315 billion for personal data breach in its 2025 complaint loss table. These are selected classes, not an exhaustive partition of digital harm. Investment fraud often involves induced transactions; it should not be counted as proof of a network breach. [82]

A different measure of impersonation

For January 2025 through July 2026, the FBI later reported nearly 61,000 complaints of law enforcement or government impersonation scams and over $1.6 billion in losses. That multi-period PSA is not an additional 2025 loss category that can be added to the annual report. [108]

Defend the business process

Treat a message, human-sounding voice or familiar title as a claim, not proof of authority. Confirm payment changes in a separate, trusted channel using an established contact record. Require two independent reviewers for material payment changes and protect the identity of people who pause suspicious requests. A valid employee account is not alone enough to authorize new financial instructions. [107] [108]

What to record

Separate attempted fraud, approved exceptions, settlement loss, funds frozen and cash returned. Record whether the proposed change would have cleared normal control gates. Report time to first finance escalation and the share of high-risk changes subjected to independent verification. Process evidence is more reliable than counting how many employees viewed training.

Exhibit 07 · Large loss categories are not incident counts

Selected crime types from the 2025 IC3 loss table. Category definitions and reporting behavior affect comparisons. These losses are not national totals, and the categories shown are not the entire complaint portfolio.

FBI, 2025 IC3 Annual Report: primary source

The AI descriptor is neither attribution nor causation

Reference to AI · 22,364 tagged complaints · Avoid causal subtraction

What is in the record

IC3 documented 22,364 complaints with its AI-related descriptor, associated with $893,346,472 in adjusted losses in 2025. Appendix C describes the label as a reference to AI in reported information. It is supplementary to a crime type. It should not be summed with other FBI loss categories as if it were another disjoint bin. [82]

What is not in the record

The label does not establish that an AI model was used, that the complainant correctly identified its role, or that the conduct would not have occurred without AI. Nor does absence of the label rule out AI use. A loss associated with a tagged complaint is not the loss caused incrementally by AI.

Different sources can corroborate different propositions

FBI has warned of a specific campaign involving AI-generated voice messages. Public reporting of such observations strengthens confidence that AI is being used in some cases, not that its attributable share is known. Separately, the March 2026 intelligence assessment described potential improvements to the speed and effectiveness of both malicious operators and defenders. An intelligence forecast is not population measurement. [107] [91]

Measure locally

Use an internal evidence ladder: suspected, independent evidence of AI participation, and demonstrated material difference to outcome. Preserve communications subject to privacy and retention policies, and treat any separately recorded model involvement as an evidentiary attribute rather than a motive or verdict. Update classification when new evidence justifies it.

Attempts, freezes and money returned

Rapid financial intervention · Selected program denominator · Freeze is not restitution

The 2025 asset recovery subset

For its 2025 fraud-intervention program, the FBI reported 3,900 incidents, $1.164 billion in attempted theft and $679.0 million in funds frozen. The annual report presents an approximately 58% success rate corresponding to dollars frozen relative to attempted theft within this pathway. It is not the proportion of all victims made whole. The subset involved both domestic and international cases. [82]

Why time matters

A financial institution and relevant investigators can sometimes intervene before money is transferred further or lost. Fast notification creates a chance for intervention without establishing that intervention will succeed. Maintain verified contact channels and backup authority in the finance response plan.

Record distinct court and balance sheet milestones

A freeze restrains a transaction or funds; subsequent release, forfeiture, recovery or restitution are different legal and financial outcomes. A service provider’s favorable press update should not be booked as cash recovered. Avoid reporting estimates from an active investigation as final reimbursement. [82] [101]

Prepare practical response authority

Define who can notify the financial institution, who can authorize emergency transaction review, and who can secure the incident timeline. Record the amount and time of each stage once verified. If the organization does not know the current disposition, state that explicitly. Testing the response handoff in advance complements technology safeguards and reduces confusion without promising recovery.

Exhibit 08 · Frozen funds are not money returned

The FBI reports about $1.164 billion in attempted theft and about $679.0 million frozen in 3,900 incidents in its 2025 targeted fund-freeze process. Amounts frozen are not necessarily returned and are not a population recovery rate.

FBI, 2025 IC3 Annual Report: primary source

[*REDACTED*]’s breach sample is a different lens

Contributed breach records · 31% initial entry classification · 48% third-party involvement

The 2026 edition

[*REDACTED*]’s May 19 release for its 19th Data Breach Investigations Report, based on 2025 data, reports that 31% of breaches began with vulnerability exploitation in its corpus, surpassing stolen credentials as the leading initial entry category in that edition. It also reports third-party involvement in 48% of breaches and describes that measure as up 60% relative to its prior comparison. [84] [85]

Statistical scope

A contributed case corpus reflects sector mix, detection and investigative access. Treat the percentages as sample observations, not the probability a randomly selected organization suffers a breach. Third-party involvement describes a feature across cases, not a separately additive initial entry route.

Interpretation for a security leader

Maintain an accountable asset and application inventory, prioritize remediation by credible evidence and business context, and require vendor escalation and recovery evidence. Do not abandon credential protections because another category passed them in one edition. Instead, compare your own incident and exposure records against the controls that could have made a difference. [97] [98]

Do not promote association into proof

A source’s narrative may connect AI with speed, but the aggregate 31% share does not isolate an independent effect of AI. A dashboard should retain the source edition, observation year, defined population, missing-data policy and whether the number is a share, comparison or experiment.

Exhibit 09 · Breach measures are not mutually exclusive

[*REDACTED*]’s 2026 report analyzes more than 22,000 confirmed breaches from October 2024 through November 2025. These three measurements describe different attributes and can overlap. Do not sum them or generalize the contributor sample as a census.

[*REDACTED*], 2026 Data Breach Investigations Report: primary source

Ransomware measures do not share one denominator

Bank Secrecy Act reports · Industry cost exceeds ransom · Recovery is operational

The FinCEN window

FinCEN’s December 2025 analysis uses Bank Secrecy Act reports filed through February 1, 2025 for ransomware incidents occurring in 2022–2024. It covers 7,395 reports tied to 4,194 incidents and more than $2.1 billion in reported ransomware payments. Its notes include actual and attempted transactions in the analysis. The agency reports about $1.1 billion and 1,512 incidents for 2023 and about $734 million and 1,476 incidents for 2024. [86]

Why the numbers differ

IC3’s 2025 complaint-loss table lists about $32.3 million for ransomware, a separate victim-reporting population and a different period. Neither figure captures the full economic harm; compare time windows and definitions before making any rate, trend or multiplier assertion. [82]

Budget response costs explicitly

Count restoration, emergency staffing, lost transactions or capacity, reconciliation, legal and notification costs, downtime, and residual remediation only where observable. Keep transfers, insurance and recovered funds separate from direct physical or operational harm. Consequences can persist after contact with criminals ends.

What makes recovery credible

A continuity plan should prioritize safe service, identify dependencies needed to restore, and exercise the ability to rebuild from validated records with ordinary systems unavailable. The NIST model treats response as a function of broader risk management, not a one-time technical effort. [96]

Disruption can help without ending extortion

Operational action versus adjudication · International coordination · Measure durable effects

LockBit and institutional response

In February 2024, DOJ announced a disruption of LockBit with the United Kingdom and other partners. The release reported more than 2,000 targeted victims and disclosed contemporaneous court actions. Later releases include particular guilty pleas. A dated enforcement action and an individual plea provide stronger evidence of their own scope than a generic assertion all members were caught. [88] [89]

FinCEN’s observed decline is not a causal experiment

FinCEN links a decline in BSA-reported incidents and payments from 2023 to 2024 temporally to disruptions of high-profile groups. The public figures alone cannot isolate how much changed because of enforcement, reporting delays, refusal to pay, or shifts to other groups. [86]

A letterhead change proves little

Extortion operations can diminish, reappear or move among multiple groups. For a victim organization, the practical question is whether integrity, access, records and services are back within agreed tolerances. A public claim that attackers departed is weaker than evidence from restored service and independent review.

No negotiating playbook

Prepare who coordinates continuity, legal decisions, finance, external reporting and victim support. Treasury warns that facilitating payments can raise sanctions risks. Payment is not a substitute for evidence of sound restoration and should not be treated as routine. Seek qualified legal and incident response guidance and use established law enforcement channels. [87] [96]

Attribution is not the same as reading an IP address

Threat actor · Sponsor · Confidence and scope

What the governments assess

The March 2026 U.S. annual threat assessment identifies ongoing concerns tied to China, Russia, Iran, North Korea and non-state ransomware operators. The authorities describe different motivating objectives, including intelligence, possible future disruption and finance. Canada’s assessment separately discusses criminal and state-backed activity in its national context. An official assessment is attributed judgment, not proof every incident in a category has the same sponsor. [90] [91] [92]

Read claims carefully

Record whether a public statement refers to observed activity, inferred intent, a suspected association, a court allegation or a verdict. Government sources and security vendors may name overlapping clusters differently. Shared infrastructure or patterns alone do not establish command, citizenship, direction or motive.

Defend according to consequence

Different attribution conclusions may change threat prioritization and the authorities to contact, but responders need not wait for geopolitical certainty before protecting people, preserving evidence and reducing harmful access. Separate the technical incident commander’s work from the actors who formally assess sponsorship. [96] [101]

A useful confidence line

Report who made the attribution, whether more than one authority supports it, what was directly observed and what remains unknown. In board materials, list the estimated consequences and feasible defensive actions even if the actor name remains disputed. Revisit previous assessments when qualified investigators publish additional evidence.

Critical infrastructure makes cyber incidents physical

Public service continuity · Separate preparation from impact · Operations staff own safety

Modern public assessments

CISA and partners warned in 2024 of state-sponsored activity indicating preparation for possible future disruption at U.S. critical infrastructure. An assessment of preparation is not evidence that a damaging outage occurred. In separate recent public warnings, the Canadian Cyber Centre described attempts against industrial equipment and some observed effects while distinguishing its national scale and uncertainty. It also described reported AI involvement in a broader campaign, not universal autonomous operation. [94] [93]

What safety changes

Industrial processes have hard physical operating limits. A routine office IT containment action can be unsuitable when service depends on operator visibility and equipment remains in a live process. Operators need uninterrupted means to confirm safe status and make certified operational decisions.

What to prepare

Establish accountable relationships among incident response, operations engineering, reliability and public communications. Maintain current dependency maps, backup communication, and alternate safe operating arrangements. Validate procurement, logging and lifecycle support with operational standards and human approval. [104]

How to report a rebound

Capture observed chronology and verified effects independently: actual service losses, safety advisories, hours in alternate operation, restoration milestones and possible customer harm. A system being reachable is not proof the physical process is safe or the original capacity restored. Escalate uncertainty to the qualified operations authority; do not run unapproved recovery changes.

Technology concentration can magnify both kinds of outage

Common-mode risk · Third and fourth parties · Distinguish attack from error

Infrastructure gains and exposure

Treasury identifies benefits from cloud adoption, including reliability and access for smaller firms, together with visibility, bargaining and concentration challenges. The 2025 Financial Stability Oversight Council report warns interconnected cyber incidents could have systemic effects, although it finds no significant U.S. financial-sector systemic cyber event to date. This is a vulnerability assessment, not a prediction of its timing. [100] [99]

An adverse event need not be an intrusion

The July 2024 [*REDACTED*] outage followed a faulty update, not evidence of a malicious breach. [*REDACTED*] estimated approximately 8.5 million affected Windows devices, less than 1% of Windows machines, yet disruptions spread through multiple critical services. A small affected share of a worldwide machine base can matter greatly if it includes key intermediaries. [102]

Map actual substitutability

Test whether supposedly independent backup systems share identity, management, update or communication services. Two cloud regions do not resolve a common account-wide or provider-wide dependency. Rank critical suppliers by ability to exit or continue independently, not only annual spend. [100] [101]

Insist on signed recovery tests

Invite key providers and service owners to exercise missing tools, delayed notice and partial restoration. Track whether contracts provide timely information and whether the organization can act on it. Preserve resilience practices for both malicious and nonmalicious emergencies.

Preparation makes the incident classification defensible

Event, adverse event, incident · Record confidence over time · Keep classification reversible

The NIST distinction

NIST SP 800-61 Revision 3 distinguishes an observable event from an event with negative consequences and from a cybersecurity incident concerning actual or imminent unauthorized harm. Initial events may be ambiguous. A service outage can follow defective software, authorized maintenance, malicious access, or other causes. Classify provisional facts honestly until credible evidence supports revision. [96]

A defensible evidence record

Track observation time, analyst validation time, detection path, incident declaration, key changes, evidence confidence and owners. Avoid retroactively representing a retrospectively inferred start as something that responders knew immediately.

Governance does not start at detection

NIST relates preparation and improvement to the six Cybersecurity Framework functions across the organization. Identify who can suspend a service, preserve evidence, prioritize operations and approve communications before the incident. Preserve the ability to use alternative resources if the main team’s systems are unavailable. [95] [96]

How to test the classification

Use a tabletop in which early signs have multiple reasonable causes. Observe whether the team verifies scope without assuming maliciousness, secures urgent continuity and records residual uncertainty. The standard is neither waiting for perfect evidence nor declaring a breach from every alert. Capture decisions affecting staff time or customers and revisit them independently.

Remediation is measured in exposure windows

Known exploitation is one signal · Risk and safety govern timing · Evidence of closure

Don’t confuse severity and urgency

CISA’s Known Exploited Vulnerabilities approach brings public evidence of exploitation into prioritization, helping avoid reliance on scores alone. Stakeholder-Specific Vulnerability Categorization also considers mission context and consequences. An asset has to be identified, and a control applied and proven effective, before reported closure can be trusted. [97] [98]

Prioritize safe action

Apply the appropriate vendor-supported remedy promptly according to applicable obligations and local safety/change-control rules. When changes must wait, record temporary agreed restrictions or compensating measures with owners and expiry dates. The reasoning differs between a routine office system and an operating safety-critical installation. [104]

Count real exposure time

Record first known applicability, owner notification, verified mitigation, verified repair, and reopen date if the issue returns. Report both median and tail intervals for critical assets and include unknown inventory coverage. Fast average closure can hide a smaller number of chronically unresolved high-impact systems.

Shift some burden upstream

CISA’s product guidance emphasizes defaults and authentication controls that reduce dependence on every customer manually configuring security. Compare platforms on supported updates, auditability, identity controls and recovery behavior, not just capabilities. An accepted vendor promise is an input to assurance; validate it in operation. [103] [96]

Contain, restore and validate in separate gates

Stop spread and harm · Do not destroy evidence casually · Test before reopening

Containment is not full restoration

Once a credible incident is declared, determine the most urgent people, service or information at risk. Coordinate actions across engineering, operations and business owners to prevent additional harm while retaining relevant evidence. Document why each decision was necessary and who could reverse it. Confidence in the extent of the incident should be communicated separately from confidence in any single remedy. [96]

Separate visibility and control

A team may restore one alerting or user-facing function while deeper services remain unverified. Conditions for safe restoration should cover records, permissions, connected dependencies and remaining threat indicators.

Define the before and after

Before restoration, identify trusted baselines, establish validation ownership, and agree acceptable residual risk. After restoration, check data completeness, reconciliation, security and continuity under sustained load. In systems supporting essential services, include operational constraints and partner acknowledgments. [96] [101]

Learn when not to close

Set recovery exit criteria and observe service after reopening. Keep exceptions visible and preserve improvement commitments. Record differences between restoration time against target, duration of degraded capacity, outstanding legal decisions and customer impact. This prevents a green dashboard after the first successful restart from obscuring ongoing harm.

Disclosure clocks follow different rules

Discovery is not materiality · Public notice is not root cause · Ask qualified counsel

Illustrative U.S. disclosure rule

SEC guidance generally requires a domestic registrant to file Item 1.05 of Form 8-K within four business days of determining a cyber incident is material, not simply four days after discovering it. It also says not to delay the materiality determination unreasonably. The guidance includes a limited route involving an Attorney General national security or public safety determination. [105] [106]

Do not infer universal coverage

Different regulators, contract terms and legal regimes may impose separate reporting and notification obligations. A deadline under one rule is not the deadline under all rules. The filing is a sourced claim about the issuer at that time, not an independent estimate of every incident’s effects or complete forensic history.

Maintain a written chronology

Record what happened, when it was found, who determined materiality or other reporting triggers, which obligations were considered, and how notice was authorized. Report material uncertainty honestly while protecting information needed for remediation, personal privacy or safety. Coordinate the business narrative with the verified record. [106] [96]

Test substitute communications

Responders must be able to communicate even when ordinary systems cannot be trusted. Exercise the handoff among incident operations, counsel, leadership and providers. Decide how corrected information will be issued. This chapter is research and does not determine which requirements apply to a particular organization.

A defensible dashboard follows harm and recovery

Operations and finance · No composite invented probability · Assign an owner

Measure stages separately

A useful dashboard tracks detection latency, incident declaration, containment, clean restoration, and actual service recovery as separate measures. Denominators should include business-critical asset coverage, completed restoration tests, and in-scope third parties. Record which events are validated, which dates are estimates, and whether observations are right-censored by unresolved incidents. [96] [95]

Separate sources of loss

Track disrupted service hours, verified transactions lost or delayed, payroll and response expense, legal costs where measurable, and funds actually returned. A fraud hold, a ransom figure, and operating profit lost to downtime are distinct. Keep a transparent record of assumptions for categories such as estimated lost sales. [82] [86]

Challenge attractive numbers

Ask whether lower detection counts reflect fewer events or degraded visibility, whether faster closure reflects easier cases, and whether vendor concentration makes losses correlated. Use severity bands and tail cases instead of a single average. For insurance and systemic resilience, constrain claims to the applicable coverage and actual exposure. [99] [100]

Update the control, not just the slide

Each selected indicator should map to an accountable remediation decision and a defined next review. Pair lagging evidence, such as service outage duration, with leading tests such as independently validated recovery capability. Review metrics with business and technical owners so that apparent improvement withstands a change in perspective.

Previous chapter

Next chapter

Back to the report

Browse: White Papers · Reports · Downloads

bottom of page