Summary

  • Stefan Savage is the Irwin and Joan Jacobs Professor at UC San Diego and an empirical systems-security researcher whose work spans network attack measurement, cybercrime economics, cloud systems, vehicles and other connected infrastructure.
  • Early work on probabilistic packet marking and backscatter reframed denial-of-service as a measurable systems problem, while keeping deployment, spoofing and vantage-point assumptions explicit rather than presenting either method as universal attack visibility.
  • Later studies followed spam and counterfeit-pharmaceutical markets through domains, affiliates, fulfilment and payment relationships, showing how distributed abuse can still depend on concentrated commercial chokepoints whose effectiveness must be measured after intervention.
  • Automotive-security work with large collaborator teams connected experimental compromise to vehicle architecture, disclosure, update governance and public safety; the evidence applies to specific tested systems and should not be collapsed into a single “car hacking” story.
  • Savage’s durable influence lies less in one invention than in a method: identify the mechanism, build a measurable system, preserve uncertainty, locate the institution with real leverage and test whether an intervention changes attacker economics or system behaviour.

The packet was never the whole attack

The most useful fact about an attack is often not the packet that arrives but the system that made the packet profitable. A denial-of-service campaign depends on machines, address spoofing, paths and victim responses. A spam business depends on affiliates, domains, hosting, fulfilment, payment processors and customers. A remote vehicle exploit depends on the architecture that joins externally reachable components to safety-relevant networks. Stefan Savage’s research career repeatedly turned these chains into measurable objects.

That approach matters because security incidents are easy to narrate at the wrong layer. A packet flood looks like a networking problem until the defender asks which machines send it, which source addresses can be forged, which networks carry it and which mitigation service has enough visibility to respond. Spam looks like a content problem until the business is mapped through acquisition, hosting, fulfilment and payment. A vehicle exploit looks like one software bug until the attack path crosses radios, head units, internal buses, control units, cloud services and software-update processes.

Savage’s work is not unified by one protocol or product. It is unified by an empirical question: what can be observed well enough to replace intuition with bounded evidence? Early work on probabilistic packet marking asked whether a victim could reconstruct the path of spoofed traffic without requiring every router to store every packet. Backscatter studies used replies sent towards forged source addresses to estimate denial-of-service activity from a network telescope. Later work followed commercial dependencies in spam and illicit pharmaceutical markets. Automotive studies mapped trust boundaries across connected physical systems.

Cloud research tested assumptions about shared infrastructure rather than accepting provider isolation as a given.

The word “bounded” is central. None of these instruments sees everything. Packet marking depends on deployment and attack traffic. Backscatter sees attacks that create the right class of replies. Payment-chain studies depend on observed transactions and entity resolution. Vehicle experiments apply to named models, software versions and attack chains. A strong empirical result does not remove these boundaries; it makes them clear enough that another researcher, operator, bank, manufacturer or policymaker can decide whether the evidence is sufficient for the action being considered.

Savage joined UC San Diego in 2001 and now holds the Irwin and Joan Jacobs Professorship in Computer Science and Engineering. He earned a bachelor’s degree in Applied History from Carnegie Mellon University in 1991 and completed a PhD in Computer Science at the University of Washington in 2002. His career includes the ACM Prize in Computing in 2015, a MacArthur Fellowship in 2017 and election to the National Academy of Engineering in 2023. UC San Diego reported in 2023 that his publication record exceeded 150 peer-reviewed papers.

Those honours matter as evidence of recognition, not as a substitute for attribution. The major systems and studies associated with Savage have co-authors, students, research staff and partner institutions. Practical IP traceback was written with David Wetherall, Anna Karlin and Tom Anderson. Backscatter work involved David Moore and Geoffrey Voelker among others. Cybercrime-economics research included Kirill Levchenko, Damon McCoy and wider teams. Automotive-security studies involved Tadayoshi Kohno, Stephen Checkoway and many collaborators. The accurate profile is not a sole-inventor story.

It is the record of a researcher and centre builder who helped make security claims answerable with data, then kept returning to the limitations of the data.

Traceback began with deployment constraints, not perfect attribution

The 2000 paper Practical Network Support for IP Traceback, written with Wetherall, Karlin and Anderson, addressed a structural weakness in denial-of-service defence. Attackers could forge source addresses, leaving the victim with packets that revealed little about the route they had taken. Keeping a complete packet history at every router was not realistic, and a global logging system would have introduced its own storage, privacy and deployment problems.

The proposed answer was probabilistic packet marking. Routers would occasionally write fragments of path information into passing packets. A victim receiving enough attack traffic could combine those fragments to reconstruct likely paths. The design tried to add evidence without requiring every router to remember every packet and without assuming a centrally operated global trace service.

Its importance came partly from its practicality and partly from the fact that its assumptions were visible. Reconstruction depended on packet volume, marking behaviour, path stability and the attacker’s traffic. Fragmentation and spoofing patterns affected what the victim could recover. The mechanism also required router deployment, which a paper could propose but not compel across the public internet. Even a reconstructed route would identify infrastructure along a path rather than prove the legal identity of the human or organisation behind an attack.

That boundary is easy to lose when “traceback” becomes shorthand for attribution. Network evidence can narrow an investigation, identify likely ingress points, help operators coordinate and reveal which providers are in a position to act. It does not by itself establish intent, ownership or culpability. The 2000 work is best understood as an attempt to make an opaque event progressively more answerable rather than as a universal mechanism that solved source attribution.

The distinction also anticipates a pattern in Savage’s later research. A security problem often cannot be solved by designing a theoretically complete mechanism that the rest of the world has to adopt at once. A more deployable approach identifies evidence that existing or incrementally modified infrastructure can expose. The resulting instrument may be incomplete, but its incompleteness can be analysed and its output combined with other institutional knowledge.

Traceback therefore belongs to infrastructure history because it asked where measurement responsibility should sit. Routers could contribute information. Victims could reconstruct. Operators could interpret paths. Other institutions would still have to decide what response was legitimate. The paper did not collapse those roles into one system, and that separation remains useful when security controls cross independently operated networks.

Backscatter turned unused address space into an observatory

Spoofed-source denial-of-service attacks generate replies to addresses that did not send the original request. When some of those forged addresses fall inside a large block of unused space monitored by a network telescope, the unsolicited replies become a sample of attack activity. Savage’s backscatter work with Moore, Voelker and collaborators used this side effect to estimate attack frequency, duration and victim patterns.

The method was powerful because it did not require cooperation from attacker or victim. The instrument watched traffic arriving at dark address space and inferred something about activity elsewhere. Infrastructure that appeared unused became a sensor. Instead of collecting incident reports from organisations willing to disclose them, researchers could observe a distributed phenomenon through an independent vantage point.

This was not a complete census. The method favoured attacks that used sufficiently random spoofed source addresses and caused victims to send replies that could reach the telescope. Non-spoofed botnets, attacks using different reflection patterns, traffic filtered upstream and attacks against services that did not generate the expected responses could be under-represented or invisible. Routing to the telescope and the size and location of the monitored address block shaped the estimate.

That is why the measurement was strong: not because it was bias-free, but because the bias could be stated and investigated. The instrument defined a population it could see. Researchers could compare durations, victim classes and attack intensity within that population while resisting the temptation to equate the sample with all denial-of-service activity on the internet.

Attack practice also changes the relationship between instrument and population. Modern botnets can send from real source addresses, abuse reflectors or use services and protocols that create different network side effects. A method developed around one era’s spoofing behaviour must therefore be revalidated rather than repeated as a timeless census. The adversary can alter the observation process by changing the mechanism of attack.

Network telescopes remain important because unsolicited traffic also reveals scanning, worms, misconfiguration and other forms of network behaviour. Their long-term value depends on stable routing, careful data handling and enough historical continuity to distinguish structural change from short-term noise. Once operators or policymakers rely on such observations, the telescope itself becomes a form of measurement infrastructure with access, privacy, governance and continuity requirements.

The backscatter programme also sharpened a distinction that runs through Savage’s career: visibility is created by vantage points, and every vantage point has politics. Who controls the address space? Who can access packet data? What identifiers are retained? Which victims are notified? What happens if law enforcement requests the data? A technically elegant instrument can create institutional power because it determines which events become visible enough to act upon.

Worms made time part of the security boundary

Worm research extended the empirical logic from individual attacks to propagation. A worm is not only malicious code; it is a process moving through a population of vulnerable machines under constraints created by scanning strategy, address space, software distribution and defence timing. Modelling that process makes the speed of response part of the security architecture.

The practical value is comparative. Defenders can examine how patching, filtering, rate limiting, address blacklisting or coordinated response changes propagation. The model cannot know the exact vulnerable population or every adaptation an attacker will make. It provides a way to test how sensitive an outcome is to those unknowns and how quickly a defensive mechanism has to act before the population changes beyond recovery.

This line of work helped shift security thinking towards systemic time. A vulnerability can be known while a fleet remains unpatched. An automated attack can propagate faster than ordinary change control. The defensive problem is therefore not only discovering a flaw but compressing the interval among disclosure, patch availability, testing and deployment without causing a different outage.

The same trade-off now appears in cloud images, operating-system fleets, software supply chains, network appliances, connected vehicles and managed devices. Faster patching reduces exposure while increasing the operational risk of a rushed change. Slower validation protects service stability while leaving the vulnerability window open. The organisation has to make the interval explicit rather than pretending that “patch available” and “risk removed” are the same event.

Savage’s broader availability research reinforces the point that technical mechanisms operate inside institutions. A containment design that works in a controlled experiment may fail when operators have incomplete inventory, incompatible software, asymmetric incentives or no authority over the systems generating traffic. Security becomes a scheduling and coordination problem as much as a code problem.

This makes historical worm research relevant even when the specific malware and network conditions have changed. The durable question is how quickly an adversary can exploit population structure compared with how quickly defenders can observe, decide and deploy. The answer depends on tooling, ownership, communication and risk tolerance, not merely the exploit itself.

Cybercrime economics followed money through a distributed infrastructure

The spam and counterfeit-pharmaceutical studies changed the unit of analysis. Counting unwanted messages could estimate volume, but it did not explain how the business survived. Savage and his collaborators linked domains, affiliates, hosting, fulfilment and payment relationships to reconstruct parts of the value chain. The visible infrastructure was highly distributed; the commercial dependencies could be much more concentrated.

That concentration created a different kind of intervention point. Removing a compromised server or domain might be cheap for the operator of an illicit business to reverse. Losing an acquiring bank, card-processing relationship or reliable fulfilment channel could impose a larger cost. The research helped show why security policy can sometimes work better when it targets the economics of a system rather than the most visible technical artefacts.

The finding is easy to overstate. Entity resolution across domains, merchants and payment records is uncertain. Test purchases reveal only observed transactions. Businesses can share infrastructure without being the same organisation. A payment relationship can change quickly. Some markets may depend on intermediaries that do not appear in the dataset. Research ethics and legal constraints limit what investigators can collect.

The defensible conclusion is therefore narrower: in the measured markets, distributed abuse depended on commercial relationships concentrated enough to be mapped and potentially influenced. That is a stronger claim than saying the entire cybercrime economy has one payment chokepoint. It also invites the next empirical question: what happens when the chokepoint is disrupted?

This is where Savage’s method becomes intervention research rather than descriptive mapping. A security study can identify a dependency, but the dependency matters only if changing it changes the adversary’s cost or behaviour. After a bank, processor, hosting provider or domain registrar acts, researchers need to measure substitution, duration, displacement and unintended consequences. Otherwise a temporary decline can be mistaken for durable success.

The shift also redistributes responsibility. Network defenders may be able to identify malicious traffic but lack leverage over the business model. Banks and card networks may have leverage over settlement but need merchant-level evidence that is appropriate to their legal authority. Regulators and law-enforcement bodies can act in ways researchers cannot. The research becomes valuable when it builds an evidentiary bridge between the mechanism and the institution actually capable of changing it.

A payment chokepoint matters only if substitution is expensive

The phrase “chokepoint” sounds more permanent than the evidence usually allows. In practice, a chokepoint is a dependency that is scarce enough, costly enough or slow enough to replace that intervention changes the attacker’s operating economics. If an illicit merchant loses one bank and opens another account the same day with no meaningful friction, the intervention did not remove a structural dependency. It merely caused a routing change in the business.

This is why pre/post measurement matters. Researchers can examine transaction success, observable sales, domain churn, pricing, campaign volume or migration to other intermediaries. A fall in one channel may represent lower demand, successful disruption, displacement to another processor or a measurement gap introduced by the intervention itself. Causal inference becomes harder precisely because the intervention changes the system being observed.

Alternative payments complicate the picture further. Cryptocurrency, intermediated digital wallets, account-based payment systems and new settlement channels can weaken some dependencies on conventional acquiring banks while creating other dependencies around exchanges, off-ramps, identity systems and liquidity. The old chokepoint may disappear; the system rarely becomes dependency-free.

That observation matters for policy because interventions have collateral effects. A payment processor, hosting company or domain provider may serve legitimate users alongside abusive ones. A network block may affect unrelated traffic. A merchant identity may be uncertain. The party with operational leverage therefore needs evidence suitable for the reversibility and severity of the action.

Temporary monitoring, rate limiting or an additional verification step can be justified at a lower confidence threshold than public accusation, irreversible account termination or legal sanction. This is not a weakness in empirical security. It is an acknowledgement that measurement and governance have to fit each other. A bounded result can support a bounded response while further measurement tests whether escalation is justified.

The deeper contribution of cybercrime-economics research is therefore not the claim that banks can “solve spam”. It is the reframing of criminal markets as systems with measurable dependencies and substitutable components. Security policy becomes a question of elasticity: how much cost can defenders impose at one dependency before the attacker moves, and how expensive is the move?

Vehicle security made software architecture a public-safety issue

The automotive-security programme with Tadayoshi Kohno, Stephen Checkoway, Karl Koscher and many collaborators attracted public attention because the demonstrations were concrete. Modern vehicles contained internal networks, electronic control units and external interfaces whose trust assumptions had developed in different engineering domains. The research showed that compromise could move across those boundaries and reach functions far beyond an entertainment system.

Reducing the work to one “car hacking” episode misses the system contribution. The programme mapped architecture, demonstrated attack paths, coordinated disclosure and forced a wider discussion about lifecycle security in machines that remain in service for years. It linked software update, diagnostics, wireless access, supplier boundaries and safety consequences.

The experiments were bounded to specific vehicles, components, software versions and test conditions. They did not establish that every model shared the same path or that researchers could remotely control any vehicle at will. Industry response also involved manufacturers, suppliers, regulators, later researchers and internal security teams. Collective credit matters because the vulnerability was a property of an ecosystem rather than one line of code.

The operational significance is that connectivity converts a vehicle into part of digital infrastructure. Wireless interfaces, cloud services, telematics, mobile networks, diagnostics, update systems and internal buses form a chain. An exploit can cross from consumer-facing software towards safety-relevant functions if trust boundaries are weak. A fix may require not only a patch but changes in segmentation, update governance, supplier contracts, monitoring and recall processes.

This also changes the economics of software lifecycle. Vehicles can remain deployed far longer than ordinary consumer software. Hardware replacement is expensive. Owners may not update promptly. Suppliers may maintain components on different schedules. A vulnerability can persist across a fleet even after a manufacturer has a technical correction ready.

Responsible disclosure is therefore inseparable from the engineering result. Publishing too early can expose owners before remediation reaches the fleet. Delaying too long can leave a serious risk hidden. Researchers and manufacturers must negotiate enough detail and time to make response possible while preserving the public value of the finding. The tension cannot be eliminated by a single disclosure rule because severity, exploitability and update capability vary.

The vehicle programme anticipated a broader cyber-physical shift. Connected medical, industrial and aviation systems face similar questions: external connectivity crosses into physical processes, software lives longer than web services, suppliers divide responsibility and safety consequences limit the tolerance for experimentation. Savage’s current 2026 engagement with vehicle-security research continues this line, but any claim about present systems should remain tied to the specific platforms and projects documented at the time.

Cloud side channels tested shared-infrastructure assumptions

Cloud computing creates a different kind of security boundary. Customers share hardware, networks and control systems while expecting isolation strong enough that one tenant cannot infer or influence another. A provider can describe the isolation model, but empirical security asks whether observable behaviour actually supports it.

Savage’s work in cloud and side-channel research belongs to this tradition of testing shared-infrastructure assumptions rather than accepting them as architectural promises. Co-residency signals, placement behaviour or hardware leakage can turn apparently separate tenants into an empirical question. The exact mechanisms evolve as providers change schedulers, processors, virtualisation layers and mitigations.

That rate of change makes historical caution especially important. A side channel demonstrated against one cloud generation may be mitigated or irrelevant in a later architecture. Conversely, a new accelerator, memory subsystem or placement policy may create a different leakage path. The durable contribution is the method of testing isolation in the environment where it is claimed rather than extrapolating indefinitely from one experiment.

Cloud work also connects Savage’s research back to infrastructure economics. Shared systems are valuable because multiplexing improves utilisation and lowers the cost of providing compute. Isolation mechanisms make that sharing acceptable. If the security boundary weakens, the economic advantage can be undermined by customers demanding dedicated resources, additional controls or separate environments.

The measurement challenge is again one of visibility. A tenant sees only part of the scheduler and hardware. The provider sees more operational data but may not expose it publicly. Researchers may have to infer placement or leakage indirectly. A result can therefore be technically correct and still have uncertain population-level significance if the fraction of affected infrastructure is unknown.

This is a familiar Savage pattern: the instrument creates a bounded view, and the interpretation must stay aligned with the bound. A demonstrated mechanism is strong evidence that an assumption can fail under named conditions. It is not a census of every provider, region or hardware generation. The next question is whether the provider can detect, mitigate and verify the issue at fleet scale.

Measurement changes the politics of who is asked to act

A measurement does more than describe a threat. It can redirect attention and responsibility. When denial-of-service is counted through backscatter, network operators and mitigation providers receive evidence about attack scale. When spam campaigns are mapped through payment relationships, banks and card networks become security actors. When vehicle interfaces are shown to cross safety boundaries, manufacturers, suppliers and regulators face a problem that cannot be left to an ordinary IT team.

This redistribution of responsibility is one reason empirical work has policy consequences. Technical communities often focus on the actor closest to the packet or vulnerable code. A system map can show that another actor has more leverage. That finding can be uncomfortable because the intermediary did not create the abuse and may serve many legitimate customers. Intervention shifts cost as well as risk.

The party asked to act needs evidence appropriate to its authority. A network operator may need prefixes, paths and traffic signatures. A bank needs merchant and transaction evidence. A manufacturer needs a reproducible vulnerability and affected architecture. A regulator needs a record that supports proportionate action. Research that identifies a chokepoint without preparing the evidentiary bridge to the responsible institution may create a headline without a practical intervention.

Savage’s body of work repeatedly built that bridge, although the outcomes differ by domain. Some mechanisms changed how a field measured a problem. Some influenced disclosure or industry practice. Some exposed intervention points whose long-term effect was difficult to isolate from adaptation. The accurate profile keeps these categories separate rather than converting influence into a universal claim of policy success.

There is also a risk in the other direction. Institutions with authority may demand impossible certainty where a reversible defensive step would be justified. Security decisions rarely have the luxury of perfect attribution. A mature process separates reversible and irreversible actions, states the confidence behind each and measures the result after deployment.

That discipline protects both sides of the decision. It prevents a bounded measurement from being stretched into a punitive conclusion it cannot support, and it prevents uncertainty from becoming an excuse for inaction when a limited defensive measure is proportionate. Measurement gains authority by fitting its claim to the decision rather than by claiming omniscience.

Research ethics become harder when the subject did not volunteer

Internet security measurement often observes people and systems that did not join a study. A telescope receives traffic from attackers, victims and misconfigured devices. A spam investigation may make test purchases and collect commercial records. Vehicle research can expose a flaw whose disclosure affects owners and manufacturers. The ethical question is not an appendix; it shapes what can be measured, retained, published and acted upon.

Researchers need to minimise harm, protect sensitive data, coordinate disclosure and avoid turning a measurement into a new attack surface. They also need to describe uncertainty so that policymakers or companies do not treat an inferred entity relationship as a proven identity. The more actionable a dataset becomes, the more consequential a false positive can be.

The same issue appears in intervention research. Identifying a payment chokepoint may encourage action against an intermediary that also serves legitimate commerce. Measuring a vulnerability at scale may expose systems before owners can patch them. Publicity can accelerate remediation or teach adversaries. Ethical review therefore has to consider the mechanism of response, not only whether the underlying data were technically accessible.

Responsible disclosure is one practical bridge. Researchers can coordinate with vendors and institutions before publishing high-impact flaws, allowing remediation while preserving the ability to document the finding. The process remains contested because vendors may ask for more time, researchers may fear indefinite delay, legal risk may affect publication and incomplete fixes can leave users exposed.

Measurement ethics also covers collateral traffic and data retention. A network telescope may receive identifiers or payload fragments the researchers did not seek. Payment studies can touch real customer transactions. Vehicle experiments may require hardware that is expensive or scarce. The instrument has to be designed not only for technical validity but for accountable access.

Savage’s work is most useful when read with these limits intact. Measurement creates leverage because it turns anecdotes into evidence. It also creates power for the party that defines the sample, labels the actors and chooses the intervention. A mature empirical-security programme makes that power visible, documents uncertainty and evaluates what happened after others acted on the result.

UC San Diego became part of the measurement infrastructure

Savage’s institutional role is tied closely to UC San Diego and the Center for Networked Systems. Official pages have used both director and co-director wording for the centre at different times, so the exact current title should be dated rather than silently harmonised. The more stable fact is that UC San Diego has been his academic base since 2001 and that centre leadership forms part of the environment through which his research is organised.

A research centre is infrastructure in a different sense from a router or telescope. It assembles faculty, students, engineers, industry relationships, equipment, datasets and administrative support long enough for projects to move from one paper to a sustained programme. Large empirical studies often depend on unusual access that a single researcher cannot maintain alone.

That structure helps explain why collaborative authorship is not a courtesy detail. Students and co-authors design mechanisms, write code, run experiments, negotiate data access and interpret results. Industry partners may provide systems or disclosure channels. External institutions can supply measurement vantage points. The centre coordinates these resources without making every result attributable to its senior leader.

Centre leadership also changes the kind of influence a researcher can have. Agenda setting, mentorship and industrial collaboration can shape which questions become tractable and which datasets become available. A senior researcher may influence a field by building the institution that allows repeated measurements, not only by publishing the most cited paper.

This institutional effect is difficult to quantify. Publication counts and awards show output and recognition. They do not reveal which infrastructure costs were paid by grants, which access came through collaborators or how many failed experiments were necessary before a public result. The source material does not support a consolidated laboratory budget or a complete map of funding.

The absence of a clean financial number is not a defect in the profile. University research is funded through appointments, federal and foundation grants, centre partnerships, industry support and project-specific arrangements. The relevant editorial discipline is to identify the type of support when documented and avoid converting collaboration into ownership or a claim that a sponsor controlled the result.

Awards recognise a programme, not sole ownership

Savage received the ACM Prize in Computing in 2015, was named a MacArthur Fellow in 2017 and was elected to the National Academy of Engineering in 2023. UC San Diego reported more than 150 peer-reviewed papers in a 2023 institutional account. These are strong signals of sustained influence across systems and security.

They do not settle attribution within multi-author research. Traceback, backscatter, payment intervention and automotive-security work each involved collaborators whose ideas, code, measurements and institutional support were necessary. Students and research staff often built the systems that made a large study possible. Centres supplied equipment, access and continuity.

This distinction is more than etiquette. Infrastructure knowledge survives when methods, datasets and institutional practices outlast one researcher. A centre that trains students and maintains measurement systems can produce a lineage of work. A profile that concentrates every achievement in the senior figure obscures the mechanism by which the field advances.

It also creates technical error. A reader who believes one person “invented” a programme may assume that person controls all implementations, datasets or later extensions. In reality, co-authors move to other institutions, datasets gain new stewards and follow-on work changes the method. Collaborative attribution gives a more accurate map of where knowledge and decision rights reside.

Savage’s awards therefore belong in the profile as recognition of a broad programme. They are not technical evidence that a specific mechanism is correct, not proof that a given intervention caused a policy outcome and not evidence of company ownership or personal wealth. The source material does not verify a general commercial ownership claim and does not support financial estimates of personal net worth or equity.

The portfolio is broader than cybercrime economics

Savage is often associated with empirical cybercrime measurement, but the portfolio extends across networking, systems security, availability, cloud infrastructure and cyber-physical systems. That breadth matters because the methods do not all produce the same kind of output or authority.

IP traceback research developed protocol and algorithm mechanisms intended to help defenders reconstruct likely attack paths from packet evidence. Backscatter measurement used network telescopes to estimate attack activity. Worm research examined propagation and containment. Spam and e-crime studies mapped commercial dependencies. Payment intervention research asked whether economic chokepoints could change behaviour.

Automotive work tested vehicle architecture and remote attack chains under controlled conditions. Cloud research examined co-residency and shared-infrastructure leakage. Disclosure work translated technical evidence into remediation processes. Teaching and mentorship extended the programme into human-capital infrastructure through students who carried methods into other organisations.

Each activity has a different practical limit. Traceback never became a universal global service. Backscatter has sampling bias. Worm models depend on historical network conditions and uncertain vulnerable populations. Cybercrime markets adapt. Payment systems can be substituted. Vehicle findings are model- and version-specific. Cloud providers change schedulers, hardware and mitigations.

The unifying asset is not a specification or commercial product. It is a way of turning hidden dependencies into observable mechanisms. This makes Savage difficult to compare with a single “competitor”. CAIDA and other measurement groups operate related internet-observation infrastructure. Security-economics researchers study incentives and markets. Commercial threat-intelligence companies collect proprietary abuse data. Automotive-security labs test other manufacturers and architectures. Law-enforcement bodies hold investigative authority and evidence unavailable to academics.

These groups overlap with Savage’s work without being substitutes. A commercial threat-intelligence provider may have current customer telemetry unavailable to researchers, while its methods and datasets may be less open. A university telescope can produce reproducible long-run evidence while seeing only one class of traffic. Law enforcement can compel information but may not publish methods. Academic researchers can expose architecture and causality while lacking authority to remediate directly.

The comparison clarifies the role: Savage’s distinctive contribution is the consistent search for measurable dependencies and intervention points across technical and economic systems. The approach is powerful precisely because it relies on unusual datasets, collaborative teams and an explicit willingness to say what the instrument cannot see.

Funding, ownership and geography place limits around the story

Savage’s work is funded through university appointments, research grants, centre partnerships and project-specific support. Industry-sponsored research and centre membership can provide equipment, data access and engineering relationships. Funding does not by itself validate or invalidate a finding, but the issuer, project and conflict-disclosure context should remain visible.

The available evidence does not provide a current audited personal laboratory budget. It also does not support a general company-ownership claim, a personal net-worth estimate or a consolidated financial value for the research programme. Awards may include fellowship or recognition funding, but they should not be treated as wealth evidence.

The institutional base is San Diego, California, while the systems studied are global. Attack traffic crosses countries. Hosting and payment chains span jurisdictions. Vehicle supply chains involve manufacturers and suppliers in multiple regions. Cloud infrastructure is globally distributed. The geographical meaning of the work therefore lies more in the systems and samples than in the address of the university.

A vantage point in one address block or one commercial market does not automatically represent the world. Geography becomes part of the sampling model. Payment relationships vary by country. Hosting markets have different concentration and enforcement regimes. Disclosure law differs by jurisdiction. Vehicle models and software versions are released unevenly.

The same point applies historically. Savage completed doctoral work at the University of Washington in Seattle before joining UC San Diego in 2001. Early network conditions reflected the architecture and attack practices of the period. Later work moved into commercial payment systems, cloud platforms and cyber-physical infrastructure. The continuity lies in empirical method, not in an assumption that the same dataset or intervention applies across decades.

The constraints qualification cannot remove

Vantage-point bias is the first structural limit. Every telescope, transaction feed, test platform or experimental fleet sees only part of the system. Better instrumentation can reduce uncertainty, but no vantage point becomes a complete census simply because its dataset is large.

Adversary adaptation is the second. Attackers can change protocols, providers, domains, payment methods, exploit chains and operational timing after an intervention. A result that was accurate before action may become less representative because the action changed the system.

Entity resolution is a third. Domains, affiliates, processors, campaigns and infrastructure have to be linked from noisy records. Shared services can make unrelated actors appear connected, while deliberate obfuscation can hide related actors. False links distort both market structure and the legitimacy of intervention.

Collaborative attribution is a fourth. Major results have many authors and partner institutions. Simplifying the record around a prominent senior researcher can produce a heroic narrative that is editorially attractive and technically wrong.

Disclosure risk is a fifth. Publishing an exploit can help defenders and attackers. Delaying can leave users exposed. The right timing depends on severity, patchability, fleet update capability and the willingness of affected institutions to act.

Research ethics is a sixth. Measurement can collect unsolicited traffic, involve test purchases or touch real systems. Privacy, collateral harm and informed-consent assumptions differ from ordinary laboratory experiments because the measured population may not know it is being observed.

Causal intervention claims are a seventh. Markets change for several reasons at once. A decline after a payment intervention does not establish that the intervention alone caused it, and a later rebound may reflect adaptation rather than failure of the original measurement.

Historical technology drift is an eighth. Internet architecture, botnets, payment methods, cloud hardware and vehicle software change quickly. Old work can remain methodologically important while becoming a poor description of current operational conditions.

Award overreach is a ninth. Prestigious honours summarise a career in broad language. They cannot substitute for mechanism-level attribution or current production evidence.

Current-role ambiguity is a tenth. Official UC San Diego pages have used director and co-director wording for Savage’s Center for Networked Systems role. The exact current title should be checked when material, rather than smoothed into one permanent label.

These constraints are not footnotes that weaken the programme. They are part of the programme’s central lesson. Empirical security gains credibility when the relationship among instrument, population and conclusion remains visible.

Measurement earns authority only when its blind spots remain visible

Empirical security research gains influence because it replaces speculation with observations that other institutions can use. The same influence creates a risk: a striking dataset can appear more complete than the vantage point permits. A network telescope, a set of test purchases or an experimental vehicle can reveal a mechanism with unusual clarity while leaving large parts of the population outside view.

The backscatter work illustrates the point. Unsolicited replies reaching dark address space made a category of spoofed-source attacks measurable without cooperation from attackers or victims. The inference was powerful because its assumptions could be stated. Attacks using other addressing or reflection patterns were less visible. As adversaries changed, the relationship between instrument and population changed with them.

Cybercrime supply-chain research created a different vantage point. Domains, affiliates, fulfilment services and payment relationships exposed dependencies packet counts alone could not show. Concentration in acquiring or settlement created leverage, yet identification remained probabilistic and markets could adapt. A payment intervention that reduced observed sales in one channel might displace transactions elsewhere.

Vehicle research added a disclosure problem. Demonstrating a path from an external interface to safety-relevant functions could accelerate repair and public understanding, but publication timing affected owners and manufacturers who had not volunteered for the study. The value of the work depended on controlled testing, collaborative attribution and a disclosure process that gave affected parties a practical opportunity to respond.

These examples show why intervention and measurement should be designed together. A study that identifies a chokepoint should specify the observable change expected if action works. It should also identify plausible substitutes and collateral effects. Institutions can then choose a proportionate, reversible response before escalating to irreversible action.

The authority to act remains outside the paper. Network operators, banks, manufacturers and regulators each possess different tools and legal duties. Researchers can explain a mechanism and estimate its importance, but they do not inherit the institution’s power simply because they supplied the evidence. Conversely, an institution should not demand impossible certainty when a limited defensive action is justified.

Long-term research infrastructure matters because adversaries adapt. Datasets, telescopes, experimental platforms and industry relationships must survive long enough to test whether an observed effect persists. Students and collaborators carry methods into new settings and often build the systems that make the evidence possible. A profile that concentrates every result in the senior researcher misses the institutional mechanism by which empirical security accumulates knowledge.

The lasting method is to find a narrow dependency inside a large problem

Across denial-of-service, worms, spam, payments, cloud systems and vehicles, the recurring move is to decompose a broad threat into mechanisms that can be observed. The attacker may be distributed, but traffic has paths. The market may use thousands of domains, but payments may converge. The vehicle may contain many components, but trust crosses specific interfaces. Once the dependency is visible, defenders can ask which actor has the authority to change it.

That method has limits when the system adapts faster than the measurement. Attackers learn which channels are monitored. Encryption and service consolidation can hide detail. Legal and commercial relationships change. The relevant metric can move from packet volume to account creation, from card payments to another settlement system or from one vehicle interface to a cloud service. A successful measurement programme must keep changing its instrument.

The public-interest test is concrete. Does the work improve the ability to estimate scale, explain mechanism or evaluate intervention? Does it preserve collaborators and evidence limits? Does it lead to a response whose effects are measured rather than assumed? These questions are more demanding than asking whether a paper was influential and closer to the reason empirical security matters to infrastructure operators.

The standard for future work is therefore demanding but usable. New measurements should state their vantage point, preserve collaborator credit, disclose ethical controls and define the observation that would weaken the conclusion. Interventions should be evaluated over enough time to detect substitution. Institutions that act on the evidence should retain responsibility for proportionality and error.

Savage’s durable contribution is this discipline of decomposition. A broad threat is reduced to traffic, dependencies, incentives and decision points that can be observed. The result does not make the adversary fully visible. It gives operators and policymakers a better basis for asking which institution can intervene, what action is proportionate and what evidence would show that the system changed rather than merely moved.