
Cyber Risk from Early Hacking to Agentic AI
Executive findings and evidence
Executive summary
FIELD NOTE Measure effectiveness in restored and trustworthy services, not in the number of alerts closed.
Old failures, new scale
The historical record has repeated patterns: interconnected systems failed together, trusted identities or suppliers widened exposure, response channels were not ready, and service restoration lagged containment. Faster tools make these weaknesses more expensive, but they do not erase the difference between access, impact and recovery. FBI IC3 recorded 1,008,597 complaints and $20.877 billion in reported losses for 2025. This includes reported internet-enabled fraud and is not a tally of all cyberattacks or economic damage. [1]
AI can help harmful actors increase volume and manage more work, and selected provider cases show human-directed automation. That evidence does not establish the prevalence of fully autonomous attacks or remove the need for independent verification. Capabilities vary with permissions, tools and environment. [10] [9]
What to do now
Inventory essential services, assign recovery owners, enforce phishing-resistant authentication where feasible, and test clean restoration. Put AI agents behind a separate permission boundary. Require independent approval for consequential actions and ensure workflows stop when approval, policy or logging fail. Avoid placing the same identity or provider on both sides of a critical control. [3] [5] [8]
Quantum computing creates a distinct long-horizon risk for widely used public-key systems. The practical response is inventory, risk prioritization and planned migration using current standards. A standards deadline is not a scientific forecast; fault-tolerant quantum capability does not itself imply AGI. Treat AGI, superintelligence and a singularity as different hypotheses. This report separates established history, measured current capabilities, and conditional future scenarios. [11]
A common lens for comparing incidents
FIELD NOTE Use the same questions for each case, but preserve what source records cannot tell us.
The six questions
What service mattered? What trust or control assumption failed? What first signal was observable and by whom? What limited the damage? What evidence supported restored operations? What durable changes followed? A case is incomplete if it only explains initial compromise or only identifies a patch. A forced shutdown can be an effective containment decision and still harm service availability. [3] [13]
Use a dated timeline with separate columns for event, detection, public disclosure, containment, restored priority services and verified closure. Confidence and time zone matter. Separate the first suspicious signal from the moment incident command declared the scope known.
Rebound is more than recovery time
Distinguish technical recovery, financial and operational recovery, public trust, remediation, and accountability. A restored customer portal does not mean that access has been fully invalidated, data integrity independently checked, or regulatory obligations resolved. A public apology is not an operational milestone. Record known gaps instead of reporting a single universal time to recovery. [3]
When comparing decades, look for the same structural dependency in different forms. Early networks often lacked coordinated response. Modern enterprises may have integrated services but still lack a separate communication channel or a tested identity recovery plan. The comparison guides controls, not reconstruction of historical intrusions.
Exhibit 01 · Cyber resilience is a repeating cycle
Reliability and recovery are recurring responsibilities. Readiness, detection, containment, restoration, and learning reinforce each other. Critical service owners should validate recovery against agreed business and safety requirements.
48PEGASUS Research conceptual exhibit. Not empirical data.
Who does what and under whose authority
FIELD NOTE Technical skill is not a moral category. Authorization, conduct and consequences matter.
Distinguish roles
Researchers can discover defects and improve systems under defined authorization and coordinated disclosure. Incident responders preserve evidence and restore services with authorized access. Journalists, archivists, standards bodies and courts contribute different kinds of evidence. Criminal and state-linked actors may have different motivations; neither a nickname nor a headline establishes attribution or collective responsibility. [7]
Profiles distinguish conviction, charge, admission, public assessment and unresolved claims. A court document establishes what the court decided or what a party claimed; procedural posture matters. Response accounts can omit confidential or sensitive information and should not be treated as complete technical records.
Translate the history into responsibilities
A disclosure program needs a published route, dependable acknowledgments, a qualified triage owner and a way to confirm changes. Give reporting parties a safe, clear scope, and verify consent before any intrusive activity. When a service is operated by a supplier, agree who receives reports, who takes technical action and who notifies affected users. [7] [4]
Name a service owner and continuity owner in addition to a technical security owner. Establish how they will handle speed-versus-evidence tradeoffs during an incident. Leadership approval should support, not override, a documented safety decision when operations could affect people or critical services. [3]
Capability is not the same as authority
FIELD NOTE A good benchmark score cannot authorize a system to take an action.
Current systems are shaped by their tools
A text model can propose next steps. An agent can carry out actions using the permissions and tools granted by its environment, retry, or hand work to another component. A workflow can be highly autonomous in one narrow task without general intelligence. Persistent memory, broad access and automatic retries can change risk without changing the underlying model’s benchmark score. [9] [8]
Observed assisted misuse can indicate what a particular provider saw in accounts or on its platform. It does not represent activity on systems the provider cannot observe, disclose or reliably attribute. Do not read provider case counts as statistical rates.
Authority belongs outside the model
Enforce tool access and business rules at an independent layer. Separate viewing, proposing and executing. Require a meaningful approval showing the operation, target and expected consequences, and ensure it expires before a changed action. A model can produce text claiming approval, but only the authoritative approval record can authorize a tool. [6] [8]
An evaluator can examine how a model handled contextual requests and warnings in a defined setup. A system owner must additionally prove that permissions, monitoring, containment, human oversight and rollback work in the deployment actually used. A benchmark and the effectiveness of a runtime control are separate evidence streams.
Quantum risk does not require a singularity
FIELD NOTE Act on cryptographic exposure that is visible today; treat arrival dates as uncertain.
Separate clocks
The protection lifetime of data, the time required to upgrade infrastructure, and the date of any cryptographically relevant quantum machine are different clocks. Long-lived confidential traffic may warrant early action because the data can still matter by the time computers change. Long-lived signatures, roots of trust and embedded equipment create different timing issues. NIST’s October 2026 transition guidance supplies migration milestones, not an arrival date for a disruptive machine. [11] [12]
A physical qubit count or a demonstration of error-corrected memory is not evidence of a system capable of defeating current deployed public-key systems at scale. Workload, error correction, logical capacity and sustained runtime would matter.
Keep AI and quantum claims separate
A quantum processor is a form of computing hardware for suitable tasks, not a switch that makes an AI system broadly intelligent. AI may help some aspects of quantum research and quantum hardware may eventually assist particular calculations. Either claim requires a defined practical workload and a comparison with classical baselines.
AGI, superintelligence and a singularity also are not synonyms. Put claims about breadth, competence, decision authority and recursively accelerating development in different columns. A high predicted impact with low evidence may justify reversible preparations, but not a declaration that the scenario already occurred. [8] [11]
Evidence hierarchy and limits
FIELD NOTE Incidents, legal allegations, model evaluations, and theoretical feasibility require different evidence.
Match the source to the claim
Use a first-party incident report for dates and acknowledged consequences, a regulator or court record for the official action, and a technical standard for recommendations and scope. A complaint report counts what was submitted. Data contributors and selection processes differ by provider. When source methods do not support national extrapolation or causal attribution, leave the stronger claim unresolved. [1] [2]
Model cards are first-party evidence of what a provider tested, not independent measurement of how a product behaves everywhere. The owner’s tool setup, allowed actions and sensitivity of work matter. An agent evaluation cannot establish a frequency of malicious outcomes among all organizations. [9]
Separate confidence by field
An outage can be confirmed while attacker attribution remains uncertain. A guilty plea can establish conduct in a defined case without establishing a wider campaign. A company’s stated recovery date can coexist with unknown data integrity or long-term cost. Record verified facts, claims by named sources, assessments and speculation as separate types.
This is a selected narrative review through November 19, 2026. Figures are bounded by sources; conceptual schematics are not measurements. Technical details that could facilitate intrusion or copycat activity are excluded. Readers can use the decision tools without attempting to reproduce a breach. [3] [7]
Priorities for the different readers
FIELD NOTE Direct attention to decisions whose costs and reversibility are measurable.
Boards, operators and public organizations
Boards should identify essential services and approve explicit recovery objectives. Test an extended outage with one unavailable supplier, identity service or command channel. Ask when human supervision or domain expertise could actually intervene, not only whether the policy says a human is involved. Review whether the backup depends on the same failed provider. [4] [13]
Operators should compare historical incident evidence with current procedures. Build a record of health, containment, isolation and restore events. Require a named authority for changes that can affect essential services and a way to back out or stop propagation.
Small teams, schools and individual households
Focus first on a small controls stack: reliable recovery for core accounts, distinct identities for critical functions, phishing-resistant methods where supported, independent approval of sensitive payments, a verified backup restore and a trusted reporting route. Avoid measuring preparedness by a purchased security tool alone. [5]
For families and schools, a clear contact protocol for money, urgent messages and impersonation can reduce reliance on a training game or on identifying visual flaws. Confirm unusual requests through an existing trusted route and support people who report concerns early. Education is one control, not a substitute for secure workflow design.
Browse: White Papers · Reports · Downloads
