Why Scam Prevention Cannot Stay Inside One Sector

The bank sees the payment. The social platform sees the advertisement. The telecommunications provider sees the call. The scammer sees the entire operation. That difference is becoming one of the most expensive blind spots in cybersecurity.

Imagine an Australian receiving a convincing investment advertisement featuring a well known financial commentator. The advertisement appears on a legitimate social media platform, links to a professionally designed financial news website and eventually directs the visitor to a messaging application. A supposed investment adviser begins a conversation, makes several telephone calls and persuades the victim to install a trading application. The application displays fabricated profits, encouraging an initial deposit followed by progressively larger transfers. The victim's bank eventually notices unusual payment activity and intervenes, but by then the scammer has already moved the conversation to another messaging account, replaced the original website and provided a different payment destination. The bank may successfully block one transaction. The social platform may remove the advertisement. A hosting provider may suspend the website. Each organisation can demonstrate that it acted responsibly within its own environment. Yet the underlying scam operation may remain almost entirely functional. This is the uncomfortable reality of modern fraud: a scam can survive several successful security interventions because those interventions are directed at separate fragments of the same criminal operation.

The scale of the problem makes this more than an academic concern. According to the Australian Competition and Consumer Commission's 2025 Targeting Scams Report, released in March 2026, Australians made 481,523 scam reports across the participating reporting organisations, with reported financial losses totalling AUD 2.18 billion. The figures combine information from Scamwatch, ReportCyber, the Australian Financial Crimes Exchange, IDCARE and the Australian Securities and Investments Commission. They also expose a fundamental difficulty in evaluating prevention: different organisations observe different portions of the problem, often through different reporting channels, data structures and evidential standards. Financial loss statistics tell us how much damage has been reported, but they reveal considerably less about which intervention opportunities were missed before the money moved. A scam may have generated suspicious advertisements, impersonating domains, abusive telephone activity and multiple victim reports long before a bank received the transaction that made the financial harm visible.

This is why I increasingly view scam prevention as a problem of distributed visibility and fragmented authority, rather than simply detection accuracy. The technologies required to recognise suspicious domains, analyse messages, identify impersonation and assess financial transactions already exist in various forms. What remains difficult is connecting those observations quickly enough for the organisations capable of acting upon them to intervene. A useful example of the architectural direction comes from the product ecosystem operating under the Cyberoo.AI brand, which connects three complementary capabilities: Scams.Report for explainable, multilingual verification of user submitted evidence, NothingPhishy for disrupting fraudulent websites, fake applications, social media impersonation, telephone related abuse and other external infrastructure, and MuleHunt for financial harm signals, payment context and mule risk intelligence. The relevance of this model is not the mere presence of three products. It is the attempt to connect evidence originating outside financial systems with infrastructure intervention and financial risk analysis. That relationship addresses a problem that no individual sector can solve merely by improving its own detection tools.

The Scammer Does Not Respect Your Organisational Chart

There is something almost absurd about the way the security industry divides responsibility for fraud. We classify threats according to technical environments, organisational functions and regulatory jurisdictions. A suspicious website becomes a web security problem. A deceptive advertisement becomes a platform integrity problem. A malicious telephone call becomes a telecommunications problem. A compromised account becomes an identity security problem. An unexpected bank transfer becomes a financial crime problem. These distinctions are administratively convenient, and they reflect genuine differences in legal authority and technical capability. Unfortunately, the scammer has no reason to observe them. To the attacker, advertisements, websites, telephone numbers, messaging identities and payment destinations are simply interchangeable components of a single revenue generating operation. Each component performs a particular function: attracting victims, establishing credibility, sustaining communication, creating urgency, collecting money or restoring operations after disruption. The attacker can replace one component without necessarily abandoning the others.

I would describe the resulting defensive weakness as sector boundary arbitrage. The attacker benefits from the fact that different organisations possess different information, use different thresholds for intervention and operate at different speeds. An impersonating social profile may remain active while a bank investigates a suspicious recipient. A telephone number may continue receiving calls after an associated domain has been suspended. A payment account may be reviewed without the investigating institution ever seeing the fraudulent advertisement that persuaded victims to send money. The attacker does not need every component to remain available. It only needs a sufficiently complete pathway from initial contact to financial extraction. This changes how we should assess successful disruption. Removing a malicious domain is useful, but the relevant question is whether the intervention has interrupted the pathway that converts deception into financial harm. If the attacker can substitute another domain while retaining the same messaging account, operator persona and payment instructions, the intervention has removed an asset without necessarily damaging the underlying business process.

This is not an argument against specialised security products. A bank cannot reasonably be expected to administer social media accounts, and a telecommunications provider should not independently determine whether an unfamiliar beneficiary is participating in money laundering. Specialisation remains necessary. The problem begins when specialised systems are designed as though the information they produce has no operational value beyond their own environment. A bank may possess evidence of a suspicious payment destination that could help a platform identify fraudulent advertisements. A platform may detect an impersonation campaign before the associated payment accounts develop an observable transaction history. A consumer may possess the conversation that explains why an otherwise ordinary transfer is fraudulent. These observations acquire greater preventive value when their relationships can be established and appropriate portions can reach the participants authorised to act. The objective is not to give every sector unrestricted access to everyone else's data. It is to ensure that relevant intelligence reaches the right decision maker before the opportunity to prevent harm disappears.

The Australian National Anti-Scam Centre has already demonstrated why this distinction matters. Its investment scam fusion cell, which operated from August 2023 to February 2024, brought together 43 organisations across government and industry. The initiative produced a direct referral process that resulted in more than 1,000 scam advertisements, advertorials and videos being removed, alongside 220 investment scam website takedowns and the diversion of 113 attempted calls to confirmed scam numbers towards recorded warnings. Those outcomes should not be interpreted as proof that the underlying campaigns were completely eliminated. They nevertheless illustrate something important: coordinated interventions can reach different points of the scam pathway, including places that a financial institution acting alone would struggle to influence. A more recent example emerged in July 2026, when the National Anti-Scam Centre announced a fusion cell targeting fake online gambling operations, involving banks, digital platforms, telecommunications providers, regulators, law enforcement and community organisations. The inclusion of frontline community services is especially significant because the evidence needed to understand a scam is not always held by the institutions processing its transactions.

The Missing Capability Is Not Another Database

When organisations discuss information sharing, the conversation often turns immediately to central repositories, APIs and intelligence platforms. These are necessary components, but they do not address the hardest part of the problem. A report is not automatically useful simply because another organisation can retrieve it. The recipient must understand what the evidence means, how reliable it is, which entities it describes and what action the evidence could justify. Consider a consumer reporting that an investment adviser asked them to transfer money after communicating through a messaging application. The bank may need the beneficiary details, transaction context and evidence of deception. The digital platform may need the advertising account, original creative and impersonation evidence. A telecommunications provider may need the relevant number, timestamps and evidence of abusive activity. A hosting provider may need the fraudulent domain, captured content and details of the impersonated organisation. Sending the complete raw report to every participant would be inefficient, potentially invasive and sometimes legally inappropriate. Sending only an account number or URL would preserve privacy at the expense of the context that makes the indicator actionable.

The more useful concept is evidence portability. Evidence portability describes the ability to move verified observations between organisations without losing their meaning, provenance or appropriate limits on use. In practical terms, this requires a structured evidence packet that separates original material from extracted indicators and analytical conclusions. The packet should preserve the source of each observation, the relevant timestamps, the relationships between entities, the confidence attached to each finding and any restrictions on disclosure. A payment destination observed in a confirmed scam should not be described in the same way as an account merely suspected of facilitating mule activity. A telephone number appearing in several suspicious messages may have been spoofed. Two domains sharing an IP address may belong to unrelated customers of the same hosting provider. Correlation creates investigative leads; it does not automatically establish common criminal control. This distinction becomes particularly important when intelligence is used to justify financial restrictions, account suspension or other actions that can affect legitimate individuals and businesses.

A further difficulty lies in the information lost during handoffs. A user submits a screenshot, an analyst extracts a URL, another system assigns a risk label, and an abuse team forwards the URL to a hosting provider. The original evidence may have explained the impersonation, the payment instruction and the relationship with several other accounts, but none of that context survives the transfer. I call this the handoff evidence tax: the analytical value lost whenever one system reduces a complex scam event to the minimum information required by its local workflow. An apparently efficient integration can impose a surprisingly high tax if it transfers indicators without preserving the relationships that made those indicators meaningful. The solution is not to transmit everything. It is to maintain a common case model and generate appropriate evidence views for different recipients. This also allows new information to strengthen an existing case without repeatedly reconstructing the incident from the beginning.

The underlying data model deserves more attention than it usually receives. A mature scam intelligence system should represent not only entities such as domains, applications, accounts, phone numbers and profiles, but also events and relationships. A domain redirected a victim to a messaging account at a particular time. A message instructed the victim to pay a particular beneficiary. A payment destination appeared in three separately verified reports. A replacement domain reused the same application download link as an earlier campaign. Each statement should carry source provenance, observation time, analytical confidence and any relevant uncertainty. Temporal information matters because scam infrastructure changes quickly, and an association observed last month may not remain valid today. Entity resolution also requires care: shared templates and identifiers may suggest common infrastructure, while ordinary service providers can create misleading apparent connections. The goal is a defensible reconstruction of the operation, not a graph that treats every technical similarity as proof of conspiracy.

AI Has Made Fraud More Scalable, but It Has Also Exposed a New Coordination Failure

Generative AI changes this problem in two directions. Attackers can use language models to produce convincing communications across multiple languages, generate variations of advertising material, adapt scripts to different audiences and maintain conversations that previously required more human effort. Multimodal systems can process and generate content involving text, images and audio, creating further opportunities for impersonation. None of these capabilities makes every AI generated message fraudulent, and it would be misleading to suggest that automation has removed the need for human operators from all scam activity. The more defensible conclusion is that AI can reduce the cost of producing variation. A campaign that previously reused the same poorly written message may now present different language, visual identity and persuasion strategy to different audiences, making simple template matching less reliable. Attackers can alter the surface of an operation while preserving its underlying sequence of deception and payment extraction.

This creates a problem that technical indicators alone cannot solve. Suppose two scam campaigns use different domains, hosting providers, telephone numbers and languages. One targets an English speaking investor, while another targets a Mandarin speaking investor. Their messages differ substantially at the word level, but both rely on the same unusual explanation for an advance payment, the same staged account verification procedure and a similar sequence of escalating deposits. A multilingual language model may be able to identify these semantic relationships even when conventional indicator matching finds little overlap. I call this semantic campaign continuity: the persistence of a scam's underlying deception and operational sequence despite changes in language, infrastructure and presentation. It provides a useful complement to technical correlation, particularly when dealing with campaigns that deliberately rotate their visible assets. Yet semantic similarity must remain an analytical signal rather than conclusive attribution. Common fraud narratives are widely copied, and unrelated operators may adopt comparable techniques.

The rise of autonomous AI agents makes the coordination problem still more complicated. A consumer may no longer interact personally with every external website or merchant involved in a transaction. An assistant might research suppliers, process invoices, communicate with service providers or prepare payments using connected tools. This introduces the possibility that untrusted external content influences an authorised agent before a financial institution sees the resulting transaction. Microsoft's June 2026 research on MCP tool poisoning examined how malicious tool metadata could redirect the behaviour of a financial operations agent. The illustrated attack concerned unauthorised disclosure of invoice information, rather than a documented fraudulent transfer, but the trust boundary is relevant to scam prevention: a legitimate agent can perform attacker beneficial actions when lower trust information is mistaken for operational authority. A November 2026 research preprint examining MCP tool poisoning across multiple agent frameworks further highlights the difficulty of controlling malicious instructions entering through tool descriptions and responses. Such experiments should not be mistaken for measured fraud rates in production environments, but they reveal an emerging technical surface that conventional bank transaction monitoring does not fully observe.

The implications are substantial. An agent may select a fraudulent merchant because of manipulated search results, accept false account details from a document or invoke an external tool whose instructions have been poisoned. The eventual payment can be technically authorised while the surrounding commercial context is fraudulent. A bank may see a valid transfer instruction. The AI application may see a successfully completed task. The external platform may see a routine merchant interaction. The attacker sees the manipulation connecting them. Preventing this form of fraud requires trustworthy evidence about the counterparty, the original user intention, the information used to justify a payment and the permissions exercised by the agent. Once again, the weakness exists between systems. It is not enough for the bank to authenticate the instruction, for the AI developer to secure the model or for the digital platform to remove a malicious listing after the transaction has occurred. The preventive controls must be able to recognise when a legitimate workflow is being redirected towards an illegitimate financial outcome.

The Scams Prevention Framework Is Really a Test of Operational Cooperation

Australia's Scams Prevention Framework (SPF) offers an unusually relevant regulatory response to the fragmented nature of modern scam activity. The framework establishes enforceable obligations for designated services in the banking, telecommunications and digital platform sectors to prevent, detect, report, disrupt and respond to scams, supported by governance arrangements. Its introduction is staged. Relevant regulated entities were required to hold Australian Financial Complaints Authority membership from 1 September 2026, while the ACCC states that most substantive SPF obligations commence on 31 March 2027. The framework does not make every company operating in these sectors automatically subject to identical requirements, nor does it prescribe a universal technical architecture for scam intelligence. Its application depends on the designated services, relevant legislative provisions and applicable rules and codes. That distinction matters when translating policy objectives into engineering requirements.

What makes the framework interesting is the recognition that responsibility for scam prevention cannot be assigned exclusively to the institution where the financial loss finally becomes visible. The bank may be positioned to interrupt a payment, but it may have limited access to the advertisement that initiated the deception. A digital platform may identify a fraudulent account before any money changes hands, but it cannot automatically determine whether the account is associated with a particular banking beneficiary. Telecommunications providers can act on relevant abusive communication infrastructure, yet they may not know which fraudulent investment application or payment destination is associated with the caller. The SPF creates a regulatory foundation for addressing these separate responsibilities, while Australia's National Anti-Scam Centre provides a mechanism for coordinating intelligence and practical disruption activity. The operational challenge remains the translation of those responsibilities into timely, evidence supported action.

There is an important distinction between shared responsibility and shared authority. The SPF does not mean that every participant should independently perform every intervention. Banks retain responsibilities and powers associated with financial services, while digital platforms and telecommunications providers operate under different legal and technical conditions. Hosting providers, domain registrars, application distribution services and overseas infrastructure operators may fall outside the initial designated sectors altogether. Nevertheless, their actions can be decisive in stopping a campaign. A properly designed response therefore needs an authority map: a structured understanding of which organisation can act on a particular asset, what evidence it requires, which escalation route is appropriate and what feedback is expected after the referral. This is a practical reason why cross sector prevention cannot be reduced to exchanging lists of suspicious identifiers.

The implications extend into consumer complaints and dispute resolution. In September 2026, the Australian Financial Complaints Authority and Telecommunications Industry Ombudsman announced a cooperation agreement intended to improve coordination as the SPF expands the treatment of scam related complaints across sectors. That development matters because the same incident may involve failures or disputed responsibilities in several organisations. An affected consumer should not have to understand the technical division between the social platform, telecommunications network and financial institution before receiving an appropriate response. Nor should the handling of complaints become an excuse to postpone urgent protective actions. Effective consumer redress and effective operational disruption are related, but they operate on different timescales. The first addresses harm and accountability after an incident; the second seeks to interrupt the attack before the harm becomes irreversible.

Where the Connected Response Model Becomes More Than a Product Catalogue

The anti-scam industry contains numerous specialised capabilities, and their value should not be dismissed. URL detection can provide rapid identification of malicious web infrastructure. Brand monitoring can reveal impersonating accounts, domains and content. Consumer reporting services can gather evidence unavailable to automated scanners. Takedown providers can coordinate interventions with infrastructure operators. Financial risk platforms can identify unusual transactions and suspicious beneficiaries. These are legitimate functions with different technical requirements. What matters is whether the output of one function becomes useful input to another. An excellent detector that produces no actionable evidence for the organisation capable of intervention has limited impact beyond its own boundary. A rapid takedown that does not retain campaign context may force analysts to investigate replacement infrastructure repeatedly. A transaction risk decision that ignores the circumstances in which a payment was requested may lack the information needed to identify deception.

The connected response model discussed earlier offers one possible architectural answer, though not the only one. Its verification capability can turn public reports, screenshots, suspicious links and multilingual communications into explainable evidence. Its disruption capability addresses external assets across several channels, while its financial intelligence capability examines payment context and possible mule relationships. In principle, those functions support a process in which a victim report contributes to infrastructure identification, relevant evidence reaches the organisation positioned to intervene, financial indicators inform related investigations and later observations improve detection of the same campaign. The distinction between architectural coverage and demonstrated performance is essential. Having complementary capabilities does not prove that every integration is seamless, that every takedown succeeds or that a financial loss would otherwise have occurred. Independent evaluation should examine whether evidence actually moves between functions, whether recipients can act upon it and whether the resulting intervention changes the attacker's operational capacity.

One way to test such a system is to follow a verified incident through its entire lifecycle rather than measure the features advertised by each component. Begin with a suspicious multilingual report submitted by a consumer. Can the system preserve the original material and explain why the activity is likely fraudulent? Can it extract payment destinations, telephone numbers, domains and account identities without confusing observed facts with inferred relationships? Can those indicators be associated with other reports while preserving source provenance? Can the relevant infrastructure providers receive sufficient evidence to act, and can a financial institution evaluate the payment context without treating every linked account as criminal? If the domain disappears but the operator reuses the same messaging account and payment instructions, will the replacement be recognised? Finally, can the outcome improve future verification rather than remaining in a closed case? These questions expose integration failures much more effectively than a feature comparison based on whether a platform supports URL scanning, risk scoring or takedown requests.

A useful operational framework would distinguish three dimensions of success. Evidence continuity measures whether observations retain their meaning and provenance as they move between systems. Intervention relevance measures whether intelligence reaches participants with the appropriate capability and authority to act. Pathway disruption measures whether the resulting actions materially interfere with the attacker's ability to reach victims, maintain communication and extract funds. These dimensions are related but not interchangeable. A system may preserve evidence perfectly yet route it too slowly. Another may generate rapid referrals but provide insufficient context for recipients to take defensible action. A third may achieve impressive asset removal numbers while leaving the main payment pathway available. Evaluating all three produces a more realistic picture of prevention than counting alerts, reports or takedown tickets independently.

The Real Financial Intelligence Gap Lies Before the Transaction

Financial institutions generally possess the strongest visibility into payments, beneficiary histories and account behaviour. That makes them indispensable to scam prevention, but it does not mean they possess the earliest or most informative evidence of deception. A bank may observe a customer transferring money to a recently opened account. The transaction can be unusual, but many legitimate payments also involve new recipients. The decisive context may be that the customer was recruited through a fraudulent advertisement, followed instructions from an impersonating adviser or received a fake investment dashboard showing fabricated returns. These observations often exist outside the bank's systems. The customer may have them on a telephone, in a messaging application or in a screenshot that was never submitted during the payment process.

This creates a distinction between transaction risk and deception risk. Transaction risk concerns characteristics of the financial event, including amount, frequency, destination, account history and deviations from expected behaviour. Deception risk concerns the circumstances under which the payment was requested and the information that influenced the payer's decision. Neither category is sufficient in every case. A payment may look unusual without being fraudulent, while another may look routine despite being induced by a sophisticated impersonation. Connecting the two allows financial institutions to make better informed decisions, especially where fraud involves authorised payments rather than stolen banking credentials. It also provides a clearer evidential basis for distinguishing a destination directly observed in a verified scam from an account associated only indirectly with suspicious infrastructure.

The same principle applies to mule risk intelligence. Criminal payment networks may involve individuals recruited through fraudulent employment advertisements, accounts controlled by complicit operators, compromised account holders or intermediaries whose knowledge and involvement are uncertain. A payment destination appearing repeatedly in scam reports is worth investigating, but repetition alone should not establish criminal attribution. The source of each report, the surrounding transaction context and the relationships between cases need to remain visible. An effective system must be capable of expressing uncertainty rather than flattening every financial signal into a binary label. That is not merely a matter of fairness. Poor attribution can cause legitimate accounts to be restricted while genuine criminal networks adapt around simplistic detection rules.

There is also a temporal advantage to bringing nonfinancial evidence into financial analysis. The first signal of a scam may be a suspicious advertisement or victim conversation, appearing long before enough transactions have accumulated to make a beneficiary statistically unusual. If the relevant evidence can reach appropriate financial controls before a transfer is initiated, the institution gains an intervention opportunity that transaction history alone could not provide. This is a more interesting measure of intelligence value than simply asking whether the system eventually identified a fraudulent account. The strongest evidence may be the evidence that arrived early enough to change a decision.

Scammers Replace Infrastructure. Defenders Must Preserve Memory.

One of the least appreciated features of organised scam operations is their ability to absorb the loss of individual assets. A website can disappear while the operator retains the advertisement templates, contact list, social profiles, messaging accounts and payment instructions. A telephone number can be replaced while the same impersonation narrative continues. A payment destination can change without altering the victim's understanding of the supposed investment or purchase. The operation survives because its important relationships are not limited to any particular asset. This is why replacement infrastructure monitoring deserves to be treated as part of prevention rather than an optional activity after takedown.

A mature system should preserve campaign memory. When a new domain appears, analysts should be able to compare its content, redirect patterns, contact information, application links and payment instructions with previously investigated cases. When a different bank account appears in a familiar scam narrative, the relationship should prompt contextual investigation rather than automatic criminal attribution. When the same multilingual persuasion sequence reappears through a different social account, the underlying pattern may warrant correlation even if conventional indicators do not match. Campaign memory depends on structured evidence and reliable entity resolution, but it also depends on recording unsuccessful interventions. A takedown request rejected for insufficient evidence can reveal which documentation is needed next time. A false positive can identify an overly broad correlation rule. A replacement asset discovered shortly after intervention can show which parts of the attacker's operation remained intact.

This leads to another useful concept: defensive learning asymmetry. Scam operators may learn almost immediately which domains were removed, which numbers stopped working and which payment accounts attracted scrutiny. Defenders, however, often distribute those observations among separate ticketing systems, organisations and regulatory reporting processes. The attacker learns from the response as a single operation, while the defenders learn from it as isolated incidents. Connected prevention should aim to reverse that asymmetry. Every verified case should strengthen the shared understanding of the campaign, within appropriate legal and privacy boundaries. Every intervention should produce information about what changed and what remained available. The purpose is not to accumulate the largest possible indicator database; it is to make the next attempt more expensive, less reliable and easier to recognise.

A Practical Measure of Success Is Whether the Scam Still Works

The security industry likes metrics that can be counted reliably. Reports received, malicious domains identified, accounts suspended, transactions blocked and investigation tickets closed are all useful measures of activity. They are not equivalent to prevented harm. A thousand removed domains may represent substantial disruption, or they may represent disposable infrastructure that an attacker can regenerate easily. A blocked transfer may protect one customer while leaving other victims exposed to the same operation. A detection system can be accurate and still contribute little preventive value if its findings arrive after the attacker has completed the financial stage.

I would measure a connected scam response through the attacker pathway itself. The relevant questions are whether victims can still encounter the deceptive material, whether the scammer can continue communicating with them, whether the operation retains usable payment destinations and how quickly it can restore those capabilities after intervention. This does not mean that the complete campaign state will always be observable. It means that operational evaluation should acknowledge the difference between asset level outcomes and campaign level outcomes. Where direct measurement is difficult, organisations should use carefully defined proxies, such as the time between verified evidence and intervention, the proportion of cases linked to known campaigns, the number of replacement assets subsequently identified and the proportion of usable evidence successfully routed to appropriate recipients. Metrics should be compared against clear baselines and should avoid claiming prevented losses that cannot be established.

There is an additional reason to resist simple performance percentages. A system that handles easy cases at high volume may appear more effective than one investigating difficult, previously unknown campaigns. A platform that blocks suspicious activity aggressively may report impressive disruption numbers while producing unacceptable consequences for legitimate users. A bank that intervenes only when confidence is extremely high may demonstrate fewer false positives while missing opportunities to prevent harm. The quality of prevention therefore depends on calibration: the evidence required for a warning should not necessarily be the same as the evidence required for account restriction, infrastructure suspension or a criminal referral. Different actions carry different consequences, and those differences should be reflected in decision thresholds, audit records and escalation procedures.

In my assessment, the most useful operational objective is preventive conversion: the ability to transform an observation into a justified intervention before the attacker completes the relevant stage of the scam. Preventive conversion depends on evidence quality, recipient authority, timing and feedback. It can fail even when detection succeeds. It can also improve without replacing every existing security product, provided the organisation establishes better ways to connect evidence and action. That makes it a particularly useful measure for evaluating cross sector cooperation, where no single participant owns the complete workflow.

The Hardest Part Is Deciding Who Should Act, and When

Shared intelligence introduces responsibilities as well as opportunities. A suspicious account identifier may be useful to a financial institution, but publishing it indiscriminately could expose innocent people or enable further abuse. A screenshot may establish that a consumer was deceived while also containing private messages unrelated to the investigation. A telephone number associated with scam calls may have been spoofed. An impersonated social account may belong to a genuine user whose credentials were compromised. The purpose of cross sector cooperation should not be to distribute increasingly large quantities of unfiltered personal information. It should be to produce evidence that is accurate, relevant, proportionate and actionable for each intended recipient.

This requires technical governance, not simply legal agreements. Intelligence should carry source provenance, confidence, observation time, permitted uses and correction mechanisms. Recipient specific disclosure should ensure that a hosting provider receives evidence of fraudulent content without unnecessary banking information, while a financial institution receives relevant payment context without an unrelated consumer's entire conversation history. Access should be controlled, sensitive data retained only as necessary and consequential decisions subject to appropriate review. Organisations also need procedures for withdrawing or correcting intelligence when an assessment changes. An inaccurate indicator that travels rapidly through an interconnected system can cause more harm than an inaccurate indicator confined to a single internal queue. Better connectivity must therefore be accompanied by better evidential discipline.

The governance challenge becomes more complex when automation is involved. Language models can assist with extracting entities from screenshots, interpreting multilingual conversations, summarising evidence and identifying possible relationships between reports. They should not be treated as independent sources of truth. A generated summary may omit an important qualification. An entity resolution model may merge unrelated accounts. A semantic similarity assessment may overstate a connection between campaigns. These risks do not make AI unsuitable for scam intelligence, but they reinforce the need for original evidence, confidence calibration and traceable analytical decisions. In a cross sector environment, explainability is not a decorative feature added to improve user experience. It is one of the conditions that allows an assessment to be inspected, challenged and acted upon responsibly.

The Next Anti-Scam Advantage Will Come From Connected Intelligence

The most revealing fact about modern scams is not that criminals use increasingly sophisticated technology. It is that they routinely exploit the gaps between otherwise competent defensive systems. A fraudulent advertisement, a messaging conversation, an impersonating website and a payment request can belong to one coherent operation while being investigated as four unrelated incidents. The attacker does not need to defeat every security control. Sometimes it is enough to move information and victims between environments faster than defenders can establish what connects them.

The answer is not a single universal fraud database, nor a demand that every organisation abandon its specialised tools. It is a connected operating model in which consumer evidence can be interpreted, verified, structured, shared with appropriate recipients and used for coordinated intervention. The product ecosystem operating under the Cyberoo.AI brand illustrates this principle through the complementary roles of Scams.Report in explainable multilingual verification, NothingPhishy in external infrastructure disruption and MuleHunt in payment context and mule risk intelligence. The wider lesson is independent of those particular products: scam prevention becomes more complete when verification, disruption and financial intelligence inform one another rather than operating as separate endpoints. Such integration still requires independent controls, reliable evidence, appropriate authority and measurable effectiveness. Architecture creates the possibility of coordinated prevention; operational discipline determines whether it delivers.

Australia's Scams Prevention Framework provides an important opportunity to make that approach more systematic, although legislation alone cannot solve the technical and institutional problems involved. The next phase will depend on practical capabilities: better evidence schemas, faster referral mechanisms, more reliable multilingual analysis, controlled information sharing, stronger financial context, sustained replacement monitoring and clear accountability for action. These are less glamorous than announcing another detection model, but they determine whether a useful observation becomes an intervention while the intervention can still matter.

A scammer may control the advertisement, the fake website, the messaging account and the payment destination. No defender needs to control all of those environments to make the operation harder to sustain. What defenders need is the ability to connect the evidence, understand the relationships and act through the organisations that possess the relevant authority.

The scammer sees one operation. Until defenders can see the same operation, the attacker will continue to benefit from every boundary we have drawn between our systems.