
Cyber Risk from Early Hacking to Agentic AI
Hacking history and the people behind it
What the word hacker used to mean
Culture before crime · Observe the conduct · Permission is not identity
A craft label, not a verdict
MIT’s Tech Model Railroad Club preserves a dictionary based on a 1959 version written by [*REDACTED*]. Its playful technical vocabulary illustrates one influential early meaning of hacking as a clever or unconventional way of making a system do something. It does not establish a single inventor of the word or a universal ethic for everyone later called a hacker. Over time, journalists and investigators used the same label for activities as different as instrumenting software, probing under contract and stealing data. [14]
The useful distinction is not appearance, employment, age, affiliation or self-description. Ask who authorized the activity, what purpose was declared, what limits applied, and what actually happened.
A defensible vocabulary
Refer to a person as a researcher, maintainer, incident responder, user, operator, defendant or convicted participant only when the record supports it. Label alleged conduct as alleged. A research title is neither blanket permission nor a reason to assume misconduct. “White hat” and “black hat” may be recognizable shorthand, but they conceal how roles, motives and authorization can change across time. [33] [30]
Applied today, the same analysis helps categorize agent-driven work. A tool cannot grant itself new rights merely by producing a persuasive rationale. Responsibility stays with the organizations and people who authorize, supervise and document its use.
[*REDACTED*] and the culture of evidence
1947 artifact · Evidence over folklore · Debugging is not an attack
What the logbook establishes
A surviving 1947 Harvard Mark II logbook records an actual moth found in a machine component. The National Museum of American History attributes the artifact to the work of the Mark II team, in which [*REDACTED*] worked, but notes it was probably not [*REDACTED*]’s own book. “Bug” had already been used to describe engineering faults before electronic computers, so claiming [*REDACTED*] invented the word turns an excellent documented example into an unsupported origin story. [15]
The record concerns an equipment fault, not hostile access. It belongs here as prehistory of diagnosing computing problems, not as an attack case.
What translates to cybersecurity
Keep contemporaneous artifacts rather than reconstructing convenient stories later. Timestamp what the team observed, distinguish event from explanation and preserve alternative readings when the evidence does not establish causation. Failed equipment, mistakes, malicious action and a dependent service outage can look similar at first. The decision to investigate is not itself an attribution.
For executives, require a chronology that can be independently checked. For responders, record what was ruled out as well as what was confirmed. For users affected by an incident, separate restoration of the service from closure of the investigation. A clean logbook does not guarantee complete visibility, but it makes mistakes discoverable. [20]
[*REDACTED*], [*REDACTED*] and trust at internet scale
1976 publication · 2015 ACM award · Identity still matters
The constructive turn
[*REDACTED*] and [*REDACTED*]’s 1976 paper helped establish modern public-key cryptography and digital signatures. ACM awarded them the 2015 [*REDACTED*] Award and emphasized the contribution to secure communication over networks used by people who cannot independently arrange a shared secret for every connection. This is a history of defensive capability: it altered what internet commerce and distributed services could reasonably trust. It is not a claim that two people invented every underlying primitive or every deployed protocol. [16]
Cryptographic designs support confidentiality or authenticity under stated assumptions. They do not, alone, determine whether a user intended an action or whether an organization should authorize it.
How foundations become governance
Operate keys, certificates and identities as governed assets. Inventory which services can sign decisions or exchanges, who can revoke trust, and what happens during renewal or migration. Make independent verification possible where data or software crosses organizational boundaries.
The quantum connection is a migration problem as much as a mathematical one. Even if a replacement is ready, ownership records, vendor dependencies, compatibility and difficult-to-update devices can make transition slow. From the historical standardization lesson, anticipate deployment friction instead of waiting for confidence about the precise date of a future capability. Secure transformations need coordination, fallback and verification, not just a new algorithm label. [16] [22]
[*REDACTED*] and resilient network design
Infrastructure pioneer · Prevent failure by design · Security includes availability
Design against ordinary failure
The Internet Hall of Fame credits [*REDACTED*] with designing the spanning tree algorithm that made larger Ethernet networks practical and notes her work as an author in networking and security. That design contribution illustrates a distinction often lost when the story begins with a malicious actor: availability is shaped first by assumptions built into how the network routes and recovers. Designers can make whole classes of ordinary failure manageable before incident responders ever arrive. [17]
Networks can be disrupted by an accident, a dependency or a deliberate operation. The external symptom does not tell the responder which occurred.
Translate the design lesson to AI
Map dependencies that silently became essential: identity providers, routing, credential validation, processing services and vendor-managed control planes. Decide what must remain usable when one fails. Give critical services practical limits and safe ways to continue when automated components become unreliable.
Do not take “we have redundancy” as evidence that redundancy works. Ask whether failover depends on the same identity, provider or operator, and use controlled exercises to discover common dependencies. The objective is to bound the effect of a local fault. For autonomous systems, the analogous requirement is independent control over authorization and shutdown: capability can be distributed, but limits must still be enforceable. [17] [22]
[*REDACTED*] and the small unexplained discrepancy
1980s laboratory case · Evidence from operations · Do not chase every anomaly
The observation that changed the question
An archived U.S. computing symposium biography records that [*REDACTED*], an astronomer at Lawrence Berkeley Laboratory, followed a 75-cent computer accounting discrepancy that led to a computer espionage investigation. The amount is memorable, but it is not a loss estimate, an attacker fingerprint or evidence that a cheap reconciliation signal always detects a serious incident. Its value was that an unexpected result could be compared to an independent system of record. [18]
Distinguish the first visible signal from the eventual determination. Discovery begins as a discrepancy; independent evidence can strengthen, narrow or refute the initial hypothesis.
The modern detection lesson
Build reconciliations across authorization, observed activity, and business consequences. A change approved for a limited period should match the period and identity in the access record. A model-approved action should have a separate record of any external transaction that followed. Use errors as prompts for human review, not as standalone conclusions.
Budget for triage. A genuine exception trail helps prioritize anomalies by scope and consequence; indiscriminately logging everything can overwhelm the organization and create unnecessary privacy risk. Include useful event context and tightly govern retention and access. If an important record is absent, state that absence directly rather than declaring that nothing happened. [18] [20]
[*REDACTED*] and the aftermath of a worm
November 1988 · 1990 conviction · From incident to institutions
The event and the legal record
The FBI records that a self-propagating program released on November 2, 1988 spread across connected systems and impaired their ability to operate. Its retrospective estimates approximately 6,000 affected computers out of about 60,000 then connected; that is an estimate, not a fully reconciled public count. Operators observed failures across institutions, and investigators later identified Cornell graduate student [*REDACTED*]. The FBI reports a 1990 jury verdict, making [*REDACTED*] the first person convicted under the 1986 Computer Fraud and Abuse Act. [19]
Public shorthand sometimes calls this the first Internet attack. “An early major Internet-wide incident” is better supported than an absolute claim of historical priority.
The rebound that mattered
The lasting change was an institutional one. After the incident, DARPA asked Carnegie Mellon’s Software Engineering Institute to establish a computer emergency response capability. The episode exposed how hard it was for operations staff to collaborate under network-wide stress even when each site had technical expertise. [20] [22]
Scale matters independently of intent. Once a system autonomously creates work for many machines at once, its effect can exceed forecasts drawn from a single trial. Today that lesson supports conservative access and rate limits, bounded deployment, separate monitoring and a shutdown mechanism that does not assume the automated system remains cooperative.
[*REDACTED*] and incident-response institutions
1988 SEI CERT/CC 1990 FIRST · Coordination without command
A neutral coordination role
In a retrospective interview, CERT leader [*REDACTED*] described early CERT as a service organization without the authority to direct every network owner. Its value lay in making separate operators more effective. Carnegie Mellon describes the center’s role in sharing remediation information and serving as a neutral intermediary among researchers and organizations. The Forum of Incident Response and Security Teams traces its own founding in 1990 to coordination problems that became visible in the early network incidents of 1988 and 1989. [21] [20] [22]
The institutional move was from a few local technical contacts toward persistent, identifiable teams with defined responsibilities.
What a smaller organization can copy
Decide in advance who owns technical response, who can pause service, who approves statements and who contacts the vendor or sector coordination group. Maintain independent communications and make arrangements for weekends and absences. Do not depend on a single staff member or on the service that is already unavailable.
Practise exchanging a small, verified incident summary across teams. Separate observed conditions, confidence levels, protective actions and unanswered questions. Preserve customer privacy and tell partners what is authorized for onward sharing. Establish criteria for when to escalate uncertainty rather than waiting for agreement about every detail. Coordination is an operating discipline, not an automatic result of buying a security product. [22] [30]
[*REDACTED*], [*REDACTED*] and public accountability
1998 Senate hearing · Researchers as witnesses · Separate claims from outcomes
Technical warning becomes public testimony
The L0pht was a group of independent security researchers whose work reached the U.S. Senate in 1998. Members included [*REDACTED*] and [*REDACTED*]. The committee’s subsequent activities report preserves the hearing’s emphasis on widespread dependency and weak assurance, while [*REDACTED*]’s first-person account describes how the group worked with journalists to bring research into public discussion. Strong claims made in testimony were warnings within a hearing, not evidence that the projected event actually occurred. [23] [24]
What made this phase distinctive was the shift from fixing a local incident to challenging what public institutions expected of suppliers and networks.
Put claims through assurance review
Public testimony can make unknown risk visible, but a risk decision must still ask what was demonstrated, what was inferred, who owned corrective work and whether an independent party checked its completion. Use controlled testing and retain evidence internally as appropriate without publishing replicable harmful details.
Apply this discipline to AI vendor demonstrations. Separate a bounded evaluation from the actual product’s permissions, monitoring, and scale. A startling claim deserves investigation; it does not become a reliable operational forecast simply because it was heard in a prominent room. Leadership should be able to convert credible concerns into reviewable corrective actions, not just striking headlines. [30]
[*REDACTED*] and the problem with celebrity accounts
1995 arrest is documented · Separate prior sentence and new claims · Identity controls over mythology
Status and sourcing
A contemporaneous DOJ announcement records that agents arrested [*REDACTED*] in North Carolina on February 15, 1995 and describes a prior computer-related case for which he had received a one-year sentence. The announcement evidences that arrest, the existence of the prior sentence and the investigative officials’ account at that time. It should not be used as a verdict on allegations then pending or as a comprehensive judgment covering everything subsequently attributed to him. [25]
[*REDACTED*]’s place in popular history illustrates a reporting hazard. Repetition of a dramatic claim does not turn an allegation, an anecdote or a marketing description into a court finding.
Organizational controls over personality
The appropriate modern lesson is not to imitate or sensationalize the conduct. It is to make verification a designed step whenever someone requests access, a credential change or an unusual transfer. Technical familiarity, urgency or confidence are not identity evidence; approvals and independent confirmation should not depend on the persuasiveness of a requester.
For investigative or legal summaries, preserve event dates and cite the charge, admission, verdict or sentence that supports each factual statement. Avoid confidently assigning actions that have no public record. Staff awareness is helpful, but workflow rules and accountable approval points are what let someone stop and verify a questionable request without personal friction. [25] [33]
[*REDACTED*] and industrial-scale payment fraud
2009 guilty pleas · 2010 sentence · Measure the exposure chain
Established conduct and scale
The Justice Department records [*REDACTED*]’s 2009 guilty plea in a conspiracy involving computer intrusions against retail and financial organizations and stolen payment-card information. A later DOJ summary records his March 2010 sentence of 20 years. These outcomes establish the conduct covered by the pleas and sentence; they do not make every number or assertion from an earlier indictment a finding. DOJ’s later account of investigative work emphasizes collaboration across cases and use of computer forensics alongside conventional law-enforcement work. [26] [27]
Victims were not limited to the operator of one system. Payment records connect merchants, financial institutions and individual account holders.
Prevention and rebound across firms
Read the case as evidence of downstream exposure that accumulates when a valuable data type flows through several organizations. Limit retained sensitive data, make payment-environment boundaries reviewable and know how partners can suspend or verify risky activity. Measure affected accounts and data actually exposed separately from records present or transactions potentially affected.
Avoid the claim that a prosecution alone repaired customer harm. Institutions still had to scope exposure, coordinate notifications and protect affected consumers. Review the same dependencies when deploying automated services over financial or identity information. A faster workflow can extend the scale of a mistake if old limits on access or retention are not carried forward. [26] [27]
[*REDACTED*], [*REDACTED*] and authorized public testing
2016 pilot · 138 unique eligible findings · Scope is the safety boundary
A bounded public program
The Department of Defense’s 2016 Hack the Pentagon pilot explicitly invited eligible participants to report vulnerabilities on a limited set of public-facing websites; critical networks were not in scope. The department’s announcement names [*REDACTED*] and [*REDACTED*] among participants, describes more than 1,400 invited registrants and more than 250 people submitting reports, and identifies 138 findings deemed legitimate, unique and eligible for a bounty. These are pilot results, not a measure of every weakness or proof that a bounty program guarantees safety. [28]
The event illustrates a deliberate arrangement: permission and scope were determined before testing, and the department assigned staff to triage and remediate.
What public discovery does not replace
A bounty pays under the terms of a particular program. A disclosure policy creates a reporting path, which may exist without payments. Neither substitutes for routine engineering review, incident preparedness, or verification that a fix remains effective. Publish where reports go and what limits apply; resource triage so legitimate reports are not lost among duplicates or unrelated claims. [30] [29]
An AI-assisted report is still subject to human review and its owner’s authorization. Prefer workflows that keep the vendor in control of assessment, avoid sensitive production data and retain enough internal evidence to confirm that the safe remediation has actually been applied.
Mirai convictions and responsibility in distributed systems
2017 pleas · 2018 federal sentences · Case records, not an attack tutorial
The people and the record
DOJ records December 2017 guilty pleas by [*REDACTED*], [*REDACTED*] and [*REDACTED*] in Alaska in connection with the Mirai botnet. The September 2018 sentencing announcement describes probation, community service, restitution and ongoing cooperation after extensive assistance to the FBI. Those details describe a specific federal case, not a claim that the defendants were responsible for every incident publicly associated with the same software, variant or label. [34]
The defining operational fact is that a collection of internet-connected devices was taken over without owners’ permission and used collectively for disruption. Individual owners could face little warning while the aggregate impact fell elsewhere.
Treat the ecosystem as a unit of defense
Assign real owners to embedded and unmanaged systems and review what happens after a product leaves active support. Require safer defaults, credential management, secure maintenance, update expectations and a way to remove obsolete devices. Large customers can shift the market through purchasing requirements rather than expecting each consumer to perform expert defensive monitoring.
Cooperation after conviction may help investigators but does not erase responsibility or substitute for safer products. Likewise, an agent running in many places should not be trusted simply because each isolated instance looks low risk. Measure the combined effective authority and enforce limits outside the agent. [34] [20]
[*REDACTED*] and the value of cross-company disclosure
2021 sworn testimony record · Discovery by a victim team · Correlated dependencies
An executive account supported by a public record
In February 2021 Senate testimony, then-[*REDACTED*] chief executive [*REDACTED*] described how his organization investigated an intrusion into its own environment and connected the event to a trusted software supply chain. The testimony is an important first-party account for understanding recognition and disclosure, not an independent audit of every claim or of every customer’s exposure. It shows the significance of looking beyond the first local alert when independent evidence points to a shared supplier. [35]
A figure associated with discovery is not its sole author. Investigators, vendors, government counterparts and affected organizations each controlled different facts and different decisions.
How the response model applies now
Notify affected parties on a verified basis, distinguish suspected exposure from observed compromise, and coordinate with the entity that can disable or replace trust in a shared dependency. Document which owned services depend on externally produced software, identity, or automated decisions. If several functions rely on one protocol or vendor, apparent operational diversity may mask shared exposure.
Apply that test to model and tool providers. Contract for usable incident records, retain a way to pause integrations, support manual continuity and define the conditions for restoring access. The important measure is time to a verified protective decision across organizations, not simply how fast a vendor drafts a notice. [35] [30]
Citizen Lab and consent-based investigation
Civil society research · Consent is a design constraint · Attribution is layered
Why evidence from outside industry matters
Citizen Lab researcher [*REDACTED*]’s 2022 testimony describes work with high-risk civil society groups and credits researcher [*REDACTED*]’s role in the program’s discoveries about commercial surveillance. The group’s study with Thai partners iLaw and DigitalReach describes obtaining forensic evidence from participants who consented. These sources describe community-based research, not an exhaustive count of spyware use or independent verification of all parties implicated by any attribution. [36] [37]
People at risk are often invisible in commercial telemetry. Consent, data minimization and secure communications affect whether the investigation can help them without adding exposure.
Separate product from operator
Evidence may support identification of a tool family, linkage to a distributor or a finding about a particular device without establishing who made or approved the surveillance decision. State those propositions separately. A public report can be strong on one attribution layer while explicitly uncertain on another.
For contemporary incidents, use narrow collection approved by the people and organizations involved, preserve needed records, minimize secondary exposure, and coordinate necessary disclosures with those at risk. A report on a politically sensitive incident may be technically defensible and still cause harm if it needlessly identifies vulnerable participants. Do not equate protection of identity with a lower evidentiary standard. [36] [37]
Ransomware figures and the distinction between disruption and guilt
NetWalker sentence in 2022 · LockBit indictment in 2024 · No guilt finding implied by charges
Compare documented case stages
DOJ reports that [*REDACTED*] was sentenced in 2022 to 20 years in prison and ordered to forfeit $21.5 million in connection with NetWalker ransomware conduct. That is a specific sentencing outcome; it is not an estimate of worldwide losses or a tally of victims compensated. By contrast, a May 2024 indictment alleged [*REDACTED*] developed and administered LockBit. The indictment is an accusation, not a verdict. Convictions of separate affiliates do not establish guilt of the person named in a different case. [38] [39] [40]
The cases illustrate relationships between developers, administrators and participants. They do not imply that role labels have the same legal meaning in every proceeding.
Disruption is useful but bounded
A law-enforcement disruption may reduce immediate operational capacity and provide assistance to some victims, while remaining distinct from permanent elimination of the threat, restitution, a prosecution result, or recovery of a particular service. Organizations should still verify their own recovery and track copied data risk and continuity even if a public takedown is announced. [40]
When measuring progress, keep separate columns for infrastructure disabled, defendants charged, judicial outcomes, assistance available and recurrence. Use precise, dated public descriptions instead of assuming an identical outcome across associated investigations. Success against an organization is meaningful even when it does not end all crime in the category.
The rules of good-faith research
Authorization and care · Disclosure is a process · Policy is not universal immunity
The legal and policy boundary
The U.S. Supreme Court’s 2021 [*REDACTED*] decision interpreted one provision of the Computer Fraud and Abuse Act, distinguishing access to information outside authorized areas from an improper purpose for accessing information someone was already entitled to obtain. The 2022 DOJ charging policy directs federal prosecutors not to charge good-faith security research as defined by the policy. A court interpretation and prosecutorial guidance are distinct. Neither is a worldwide license or a substitute for clear authorization, other legal duties or professional advice. [32] [33]
Security teams should define limits and contact channels explicitly, rather than assuming participants understand informal conventions.
Turn ethics into a service
CISA’s 2020 directive requires covered federal executive agencies to publish a vulnerability disclosure policy. CERT/CC documents coordination roles, and the IETF’s 2022 informational RFC 9116 defines a machine-readable means of advertising how to contact a reporting program. These are complementary: a discoverable address is only useful if a staffed team can receive, assess and communicate about a report; neither a contact file nor a broad slogan creates permission for out-of-scope activity. [29] [30] [31]
Keep ownership, human approval and stopping rules outside automated tools. Protect the reporter and the public, document changes, and verify remediation before celebrating closure. That is the ethical connection between the people, cases and institutions in this chapter.
Exhibit 02 · Responsible disclosure assigns clear roles
A researcher, coordinator, product owner and affected users have different responsibilities. Agree on scope, milestones and communication. Public notices should help affected users protect themselves without creating unnecessary harm. Paths and timelines vary.
48PEGASUS Research conceptual exhibit. Not empirical data.
Browse: White Papers · Reports · Downloads
