Phishing is a form of social engineering in which an attacker uses a deceptive message, website, conversation, or other interaction to induce someone to disclose sensitive information or perform an action that benefits the attacker. Typical objectives include obtaining passwords, taking control of accounts, redirecting payments, and persuading recipients to install malicious software. Although historically associated with email, phishing also occurs through text messages, telephone calls, messaging applications, social networks, advertisements, and physical objects that direct people to online services.
Phishing exploits the relationship between human judgment and technical systems. A message may appear to come from a bank, employer, colleague, delivery service, or government agency, while the requested action takes place on attacker-controlled infrastructure or through a legitimate service used deceptively. Successful attacks do not necessarily require a software vulnerability: a technically secure application can still carry a fraudulent instruction, and an encrypted website can still belong to an impersonator.
Within cybersecurity, phishing is important both as a direct instrument of fraud and as an initial route into larger compromises. Its study therefore spans communication, psychology, identity management, payment controls, and network defense. The term covers a changing family of techniques rather than a single protocol or fixed attack sequence.
Definition, boundaries, and historical development
The defining feature of phishing is deception intended to elicit a consequential response under a false understanding of identity, context, or purpose. Common responses include entering credentials, providing financial details, approving an authentication request, opening a malicious attachment, or transferring money. A message can qualify as phishing even when it contains no link or attachment: an instruction to change a supplier’s bank details may be sufficient. Conversely, an unsolicited advertisement is not necessarily phishing, and an ordinary misleading claim is not automatically a cybersecurity attack. Classification depends on the deception, the requested action, and the intended benefit to the attacker.
Phishing is a subset of social engineering, a broader category that includes in-person impersonation, deceptive telephone conversations, and manipulation of organizational procedures. Its boundary with other online fraud is not perfectly standardized. Some definitions emphasize credential theft through counterfeit websites; others include payment fraud and malware delivery. Business email compromise overlaps substantially with phishing but is defined more by the misuse or impersonation of business communications to enable fraud. It may involve a stolen mailbox, a lookalike address, or an entirely fabricated identity.
The term became associated with account theft on America Online in the mid-1990s, when attackers impersonated service representatives and sought passwords or billing information. Its spelling evokes “fishing,” with messages cast out in the hope of catching victims, and also reflects the vocabulary of earlier telephone-system hacking. Claims about a single inventor or an exact first occurrence are less secure than the documented association with this period. As commercial use of the Internet expanded, impersonation of banks, payment services, and online marketplaces became prominent.
During the 2000s, phishing increasingly combined bulk email with counterfeit login pages and automated collection of stolen credentials. In the 2010s, cloud-hosted email and collaborative work systems made a single account potentially valuable as an entry point to documents, internal conversations, password resets, and business processes. Account compromise also enabled attackers to send messages from genuine addresses rather than merely imitate them. This reduced the usefulness of sender recognition as an independent safeguard.
The spread of smartphones and app-based services broadened the attack surface further. Text-message lures, mobile payment requests, QR codes, and fraudulent support conversations became established variants. Contemporary phishing therefore differs from the stereotype of an obviously misspelled email linking to an imitation bank. It may be a brief, fluent instruction embedded in a real conversation, a notification generated by a legitimate platform, or a request to authorize an application rather than disclose a password.
Persuasion, trust, and user interaction
Phishing works by making a harmful action appear appropriate within a plausible situation. Attackers commonly invoke authority, urgency, familiarity, scarcity, or anticipated loss. A warning that an account will be suspended creates pressure to act; a purported invoice presents payment as routine work; a message about a parcel invokes an expected transaction. These devices need not convince the recipient that every detail is authentic. They may succeed by creating enough plausibility that the requested action receives less scrutiny than it otherwise would.
The attack often has two linked components: a pretext and an action path. The pretext explains why something must happen, while the action path provides the link, number, attachment, reply address, or account details through which it happens. Effective pretexts fit the recipient’s current task. A document-sharing notification is more persuasive when the recipient is already collaborating on a project; a payroll message is more credible near a pay date. This contextual fit helps explain why the same message can be ineffective for most recipients but convincing to a small subset.
Trust is frequently transferred from a familiar visual or social cue to an unrelated technical object. A displayed sender name may be mistaken for proof of who controls the sending address. A company logo may be treated as evidence that a page belongs to that company. A padlock icon may be interpreted as an endorsement rather than an indication of an encrypted connection. Similarly, a message arriving through a familiar collaboration platform can inherit the platform’s reputation even when its content was supplied by an attacker. These are mismatches between what a signal actually establishes and what the user reasonably, but incorrectly, infers.
Phishing also exploits divided attention and the cost of verification. Workers may process large volumes of routine requests, use small screens that truncate addresses, or switch rapidly between applications. Technical details such as domain boundaries and identity-provider redirects are difficult to interpret without relevant knowledge. Organizational expectations can compound the problem when employees are rewarded for responding quickly or discouraged from questioning apparent authority. Susceptibility is consequently not a stable property that divides people into permanently “careful” and “careless” groups; it varies with context, workload, experience, and the quality of available interfaces.
Some campaigns use an extended conversation rather than an immediate demand. An attacker may first establish contact, ask a harmless question, and introduce the consequential request only later. A compromised account can supply prior correspondence and genuine signatures, making the interaction more credible. Other attacks deliberately provoke alarm and offer fraudulent assistance as the solution. In both cases, the relevant mechanism is not simply belief in a false statement but the construction of a decision environment in which the attacker’s preferred action appears normal, urgent, or necessary.
Delivery channels and major variants
Email remains a major phishing channel because it reaches large populations, supports rich formatting and attachments, and is central to account recovery and business administration. Bulk campaigns distribute broadly reusable messages, while targeted messages incorporate details about particular recipients or organizations. Attachments may be malicious themselves or may contain links, QR codes, or instructions that move the recipient to another service. A file hosted on a legitimate cloud platform can be used as an intermediate step, so the first visible link need not point directly to an attacker’s own domain.
Spear phishing denotes phishing directed at selected individuals or groups, usually with some tailoring to their role, interests, or relationships. The term whaling is often used for attacks aimed at senior executives or other high-value targets. These labels describe targeting rather than a distinct technical mechanism. A message addressed by name is not necessarily highly researched, and a broad campaign can still produce highly consequential account compromise. Automation and access to exposed organizational information make personalization possible at considerable scale.
Text-message phishing is commonly called smishing, while voice-based phishing is called vishing. Text messages frequently impersonate delivery services, toll operators, banks, or account-security teams. Voice attacks may involve live callers, recorded prompts, or a combination of channels, such as a call instructing the recipient to approve a request on a computer. Caller identification is not reliable proof of identity, and a displayed local or familiar number may not establish where a call originated. The short format of mobile messages can also leave little room for contextual evidence.
Instant messaging and social media support phishing through direct messages, group invitations, fraudulent support accounts, and compromised contacts. A message from a genuine friend’s account can exploit an existing relationship without requiring the attacker to create a convincing new identity. Collaboration platforms introduce related risks when external guests, shared documents, or automated notifications are interpreted as implicitly trusted. Fraudulent advertisements and search results can also direct users to impersonated login or support pages, although these routes are not always included in narrow definitions of phishing.
QR-code phishing, sometimes called quishing, places a machine-readable destination inside an image or on a physical surface. The code itself does not make the destination malicious; its significance is that it can obscure the address until scanned and may move the interaction from a managed workstation to a personal phone. Similarly, a phishing attempt can begin offline with a printed notice or letter. By contrast, pharming usually refers to redirecting users through manipulated name resolution or related technical interference. It can deliver an impersonated site without a deceptive message and is therefore related to, but not identical with, phishing.
Impersonation infrastructure and technical mechanisms
Phishing infrastructure generally connects a delivery mechanism, a plausible identity, an interaction point, and a way to receive the result. These components may be distributed across different providers and jurisdictions. The impersonated identity might be a bank, a colleague, or an internal support team; the interaction point might be a counterfeit website, a telephone conversation, or a genuine application authorization page. Some campaigns use attacker-registered domains, while others compromise existing websites or accounts. Legitimate hosting and communication services can reduce operating costs and lend superficial credibility without making the activity authorized.
Domain names are particularly important because users often rely on visual resemblance rather than structural interpretation. A lookalike domain may differ by a character, use an alternative suffix, or place a familiar brand inside a longer name. In a hypothetical address such as accounts.example.com.attacker.test, the initial words do not establish control by example.com. Internationalized names introduce additional visual complexity because characters from different writing systems can resemble one another. The Domain Name System resolves names to network resources, but successful resolution does not establish that a site represents the organization a user has in mind.
A counterfeit page can reproduce logos, layouts, error messages, and login prompts from a legitimate service. Through HTTP over Transport Layer Security, it can also provide an encrypted connection and a valid certificate for its own hostname. Such a certificate establishes a relationship between a key and a domain under the certificate’s validation rules; it does not certify the truth of the site’s commercial claims. The distinction between transport security and organizational authenticity is central to understanding why an HTTPS connection is compatible with phishing.
Some credential-theft systems simply collect submitted information. Others interpose themselves between the user and a real service, relaying interactions and attempting to obtain an authenticated session. This is often described as adversary-in-the-middle phishing. A session cookie or similar bearer token may represent an already authenticated session, making its theft different from theft of a password alone. This helps explain why certain attacks can remain effective against one-time codes: a code entered during a live deceptive interaction may be relayed while still valid.
Not all attacks require a counterfeit login page. In consent phishing, a user is deceived into granting a malicious application permissions through a legitimate authorization system, often using OAuth. The service may accurately display a permission request, while the surrounding story misrepresents why access is needed. Other attacks seek approval of an authentication notification, installation of remote-access software, or a reply containing confidential information. These variations show why phishing cannot be identified solely by looking for imitation websites or password fields.
Target selection, criminal organization, and consequences
The value of a phishing target depends on what successful deception makes possible. A consumer account may provide stored payment methods, personal correspondence, or access to other services through password recovery. A workplace identity may expose internal files, customer information, cloud administration tools, or approval workflows. An apparently low-privilege mailbox can still be valuable if it contains invoices, contact lists, or conversations that support further impersonation. The immediate victim and the eventual financial victim may therefore be different people or organizations.
Business email compromise illustrates the connection between communication and financial control. An attacker may impersonate an executive requesting an urgent payment, a supplier announcing changed bank details, or an employee seeking to redirect payroll. If a mailbox has been compromised, the attacker may exploit an existing transaction rather than invent a new one. The fraudulent instruction can then arrive within a genuine conversation and refer accurately to amounts, schedules, or counterparties. Such attacks may involve no malware and no obvious technical anomaly in the payment message itself.
Phishing can also initiate a broader intrusion. Stolen credentials may provide access to email, remote services, or cloud resources, after which an intruder seeks additional privileges and systems. A deceptive attachment or installation request may establish a foothold that supports data theft or deployment of ransomware. Phishing is only one possible initial-access method: exposed services, software vulnerabilities, and other forms of credential abuse can produce similar outcomes. An incident’s final effects do not by themselves reveal how the intrusion began.
The supporting economy can be divided among multiple actors. Some create and distribute phishing kits, some obtain or resell access to compromised accounts, and others specialize in financial fraud or later-stage intrusion. Service-based offerings may bundle templates, hosting, campaign management, or collection interfaces. This division of labor lowers the expertise required for some activities, while sophisticated targeting and persistent intrusion may still require substantial operational knowledge. Commercialized tooling also means that visually similar attacks need not come from the same individual or organization.
Consequences extend beyond the amount directly transferred. Account recovery, investigation, business interruption, notification duties, and loss of access can impose substantial costs. Exposure of confidential information can affect customers and partners who never received the original message. At the same time, reported losses require careful interpretation: fraud statistics may combine different schemes, reflect only reported cases, or record an attempted amount rather than a final unrecovered loss. Organizational impact depends on both the initial deception and the controls that limit what a compromised identity can do afterward.
Detection, filtering, and technical disruption
Phishing detection combines evidence from message transmission, sender identity, content, links, attachments, and subsequent behavior. Email gateways may evaluate the reputation of a sending system, inconsistencies between visible and underlying addresses, known malicious destinations, or suspicious file characteristics. Email authentication contributes important but bounded evidence. Sender Policy Framework checks whether a sending host is authorized for an envelope-sender domain; DomainKeys Identified Mail verifies a domain-associated signature over selected message content; DMARC adds alignment rules and a policy framework for the visible From domain.
These mechanisms reduce some forms of direct domain spoofing, but they do not establish that every authenticated message is benevolent. An attacker can authenticate mail from a domain under the attacker’s control, and a message sent through a compromised legitimate account may pass authentication. Display-name impersonation and lookalike domains likewise fall outside the simple question of whether a domain authorized transmission. Forwarding and mailing-list transformations also complicate interpretation, making authentication a component of judgment rather than a universal verdict.
Content analysis ranges from hand-written rules to machine learning models. Systems may examine linguistic patterns, message structure, visual resemblance to known brands, destination characteristics, and relationships among senders and recipients. Natural language processing can help identify requests for credentials or payments, while computer vision can compare rendered pages with legitimate login interfaces. Attachment analysis may include controlled execution in a sandbox to observe behavior rather than rely only on file signatures. Each method has limits: a harmless-looking message can precede a malicious follow-up, and a legitimate-looking page may change after initial inspection.
Detection can occur at several points in time. A gateway may block a message before delivery, a web browser may warn when a destination is opened, and a security service may reassess a link after its reputation changes. Account-monitoring systems may detect unusual sign-ins, unexpected mailbox rules, or abnormal access to documents after credentials have already been stolen. These are different defensive opportunities. Detecting a suspicious message does not imply that every copy was removed, while identifying post-compromise behavior may still prevent further harm.
Disruption involves hosting providers, domain registrars, email services, incident-response teams, and organizations whose identities are being impersonated. Removing a fraudulent page or disabling an account can shorten an attack’s useful lifetime, but attackers may replace infrastructure quickly. Automated reporting also needs safeguards against mistaken or abusive takedown requests. The practical effectiveness of filtering and disruption therefore depends not only on classifier accuracy but also on response speed, provider cooperation, coverage across channels, and the quality of evidence available to those making enforcement decisions.
Authentication and the limits of phishing resistance
Authentication establishes that a claimant satisfies a system’s requirements for an identity; it does not automatically establish that every surrounding interaction is honest. Password-based systems are vulnerable to phishing because a password can usually be entered into any text field and then reused elsewhere. Password managers can reduce this risk by associating stored credentials with particular sites, but their protection depends on correct domain matching and on users not bypassing that behavior through manual copying or other workarounds. They do not independently verify payment instructions or the legitimacy of an attachment.
Multi-factor authentication requires evidence from more than one category, such as a password and possession of a device. It substantially changes the requirements for account takeover, but implementations differ in resistance to deception. A code received by text message or generated by an authenticator can be disclosed to a counterfeit page and relayed in real time. A push approval can be authorized under a false pretext. Number matching and additional transaction context can reduce certain errors, but not all second-factor methods bind approval cryptographically to the legitimate website.
Phishing-resistant authentication addresses this problem by making the authentication result unusable at an impersonating service. WebAuthn, standardized through the web standards process, enables public-key authentication scoped to a relying party. The FIDO2 framework combines WebAuthn with protocols for communication between client devices and authenticators. In typical browser use, a credential registered for a legitimate service is not available to an unrelated lookalike domain merely because the page reproduces the same branding. The protection derives from protocol properties rather than from the user noticing a visual discrepancy.
Passkeys are credentials based on these public-key authentication mechanisms. They may be synchronized through a credential provider or remain bound to a particular device, depending on implementation. A biometric check or device PIN often authorizes local use of the private key; the biometric template is not ordinarily sent to the website as the authentication secret. A security key can provide related protection using a separate authenticator. These methods resist credential phishing, but the security of registration, synchronization, device access, and account recovery remains part of the overall system.
Phishing-resistant login is not equivalent to phishing-proof activity. A person authenticated with a passkey can still be persuaded to make an unauthorized payment, share a sensitive document, or approve excessive application permissions. Session theft from a compromised endpoint can also occur without defeating the original login protocol. Recovery channels may create alternative routes into an account if their controls are weaker than those of normal authentication. Consequently, evaluations distinguish protection against impersonated login pages from protection of the entire account lifecycle and from protection against fraudulent instructions issued after login.
Organizational controls, education, and incident response
Organizational defense treats phishing as a process risk rather than solely a message-recognition problem. A threat model identifies which identities, transactions, and information flows would be most damaging to compromise. Technical controls can then be combined with limits on privileges, separation of duties, and independent confirmation of high-consequence actions. For example, a payment workflow may require a change to supplier banking details to be confirmed through an established channel rather than through contact information supplied in the change request itself. This interrupts the fraud even when the initiating message appears credible.
Security awareness programs explain relevant cues, reporting channels, and the limits of familiar indicators. Effective content distinguishes between weak clues and stronger evidence: poor spelling may occur in phishing, but fluent writing does not establish legitimacy; an encrypted connection protects transport, but not the truth of a request. Training also has to reflect the tools employees actually use. Guidance limited to email cannot fully address attacks delivered through messaging platforms, telephone calls, cloud documents, or app-consent screens.
Simulated phishing campaigns are used to test controls and study responses, but their results require careful interpretation. A click rate is influenced by message design, timing, recipient role, and what the simulation counts as a failure. Clicking a link is not the same outcome as submitting credentials, approving access, or transferring money. Conversely, a low click rate does not prove that an organization can detect and contain an account compromise. Simulations can also distort behavior if employees become reluctant to open legitimate messages or perceive testing as punitive rather than informative.
Reporting systems connect individual observations to organizational response. A suspicious message reported by one recipient can reveal a campaign affecting many others, and reports can supply examples that automated filters missed. The usefulness of this feedback depends on accessible reporting mechanisms, rapid triage, and the ability to correlate related messages and accounts. Metrics such as time to first report, time to removal, and the number of affected identities can be more informative than a single percentage of recipients who interacted with a test.
Incident response depends on the action that occurred. Credential disclosure raises questions about unauthorized sign-ins, session persistence, reused passwords, and recovery settings. Application-consent abuse requires examination of granted permissions and token access. Suspected malware introduces endpoint investigation, while payment fraud involves financial processes as well as account security. Evidence may include message headers, authentication logs, mailbox rules, cloud audit records, and payment records. Revoking a password alone may not terminate all sessions or remove every persistence mechanism, so response procedures distinguish the stolen secret from the access and changes it enabled.
Measurement, research, and emerging changes
Phishing prevalence is difficult to measure because observations are collected at different points in the attack process. An email provider may count messages, a browser vendor may count visits to flagged pages, an abuse organization may count unique sites, and a law-enforcement agency may count complaints or reported losses. A single campaign can generate many domains and millions of messages, while one compromised account may support multiple unrelated fraud attempts. Apparent differences between published totals can therefore reflect different units of measurement rather than contradictory evidence.
Detection research faces related problems of labeling and representativeness. A dataset assembled from reported phishing pages may contain confirmed malicious examples but omit new or unreported campaigns. Legitimate examples collected from popular websites may be easier to distinguish from phishing pages than the legitimate login, document-sharing, and payment pages encountered in actual work. Duplicate templates or closely related domains can also appear in both training and evaluation data, inflating reported performance. These issues connect phishing research to data leakage, sampling bias, and distribution shift.
Operational evaluation must account for the rarity of some attacks relative to ordinary activity. A detector can achieve high overall accuracy while generating an unmanageable number of false alarms, or while missing the small subset of attacks with the greatest consequences. Precision and recall describe different aspects of this trade-off. Performance also changes over time as attackers replace domains, alter wording, use new services, and adapt to filtering. Evaluation on later data, unfamiliar campaigns, and realistic traffic is therefore distinct from testing on a randomly divided historical collection.
Generative artificial intelligence can reduce the effort required to draft fluent messages, translate lures, and vary content. Large language models can also support defensive analysis, classification, and analyst assistance. The existence of these capabilities does not by itself establish how much they increase successful victimization: outcomes depend on targeting, delivery, interaction design, and the controls surrounding the requested action. High-quality writing was already available to skilled attackers, so the change is partly one of cost, scale, and accessibility rather than the creation of an entirely new category of deception.
Voice synthesis and generated images or video introduce further impersonation possibilities, especially when a familiar voice or face is treated as sufficient authorization. Evidence about any particular incident still requires attention to how the material was produced and used; not every unusual recording is synthetic, and not every impersonation requires sophisticated generation. Research increasingly examines the interaction between content authenticity, account provenance, cryptographic identity, and business authorization. These are separate questions: a real account can issue a fraudulent request, an authentic recording can be reused misleadingly, and a technically verified identity can still be acting without authority for the transaction at hand.