Most people recognize a smishing text, the text-message version of phishing. A message says a package is stuck, an account is locked, or a payment failed, and it includes a link to fix the problem. That version is still the most common.
Attackers also run a slower version. Instead of sending a link right away, they start a conversation. They pose as an executive, a coworker, or someone from IT, get the target to reply, and build enough trust to make a request later. Often the first request is to move the conversation to a different app. The FBI (opens in new tab), the FTC (opens in new tab), OpenAI (opens in new tab), and Anthropic (opens in new tab) have all documented versions of this since 2025.
That’s why more of our text-message testing now runs as a conversation instead of a single message with a link. The shift follows from our work in voice phishing, or vishing, where a short call lets an attacker collect a detail or two to make the next one more convincing. A text thread removes the time pressure. The attacker can reply when it suits them, keep the conversation going for days, and build trust before asking for anything.
And it isn’t just SMS. The FBI’s December 2025 update (opens in new tab) on the impersonation of senior U.S. officials notes that the first text usually asks the target to move to Signal, Telegram, or WhatsApp. The app changes, but the techniques do not. That’s what we mean by “same influence, different medium.”
This post covers what the public reporting shows, why the employee is one layer of defense and not the only one, and what we think a test like this should measure.
What the public reporting shows
In May 2025, the FBI's Internet Crime Complaint Center (IC3) (opens in new tab) warned that attackers were sending text messages and AI-generated voice messages that claimed to come from senior U.S. officials. The goal was to build trust before going after the target's personal accounts. In some cases, the attacker sent a malicious link presented as a way to move the conversation to another messaging app. The December update (opens in new tab) gave more detail. First contact is usually a text message. The attacker briefly discusses a topic the target knows well, then asks to switch apps almost immediately.
The exchange is short, but it has a purpose: earn enough trust that the next request feels normal.
Investment fraud shows the slower version of the same approach. The FBI's 2025 Internet Crime Report (opens in new tab) says these scammers usually make first contact through a text, a social media post, an ad, or a dating app, then move the conversation to a different messaging platform. The report says these scammers use AI to generate thousands of conversations that look different to each target. In 2025, the FBI tagged 4,356 investment fraud complaints as AI-related, meaning the complaint referenced AI, with reported losses of more than $632 million. Total investment fraud losses were about $8.6 billion. The FBI notes that the gap likely reflects victims not realizing how much AI was involved.
OpenAI's 2025 reports on scam takedowns (opens in new tab) show how operators use the models. In the cases OpenAI disrupted, operators fed the model messages they had received from targets and asked it to write the next reply as the fake persona. They also used it to translate. OpenAI says every scam operation it banned in 2025 primarily used AI as a scaling and efficiency tool. The UN Office on Drugs and Crime (UNODC) (opens in new tab) reported in September 2025 that scam centers in Southeast Asia use models the same way: to draft, translate, and tailor messages so one operator can handle more conversations at once. The same report says the centers still want to hire native speakers, because the models handle cultural nuance poorly.
One case goes further than the others, and it should be read separately. In September 2026, Anthropic reported (opens in new tab) a set of dating apps whose chat personas were run by AI while being advertised as real people. That is a consumer scam, not an attack on employees. It does show that AI holding conversations with no person watching each one is already happening.
Links have not gone away. The FTC reported (opens in new tab) that consumers lost $470 million in 2024 to scams that started with a text message. The most reported type was a fake package delivery notice, which depends on a link. The same data includes "wrong number" texts that start as a friendly conversation and turn into a scam later. Short exchanges, long conversations, and plain links are all in use at the same time.
A testing program that only imitates one of them is only testing part of the threat.
The human is one layer
Companies invest in identity and access management because one person's decision should never be the whole defense. Behind every employee, there are processes and technical controls that are supposed to limit what an attacker can reach.
If someone opens a malicious attachment on a managed work computer, that alone should not be enough for an attacker to compromise the company. Endpoint detection and response (EDR) tools exist to catch exactly that. If someone types their password into a fake login page, that alone should not be enough to compromise their account, as long as their sign-in method cannot be phished. CISA's fact sheet (opens in new tab) is direct on this point: the widely available phishing-resistant option is FIDO/WebAuthn, the standard behind hardware security keys and passkeys.
Mandiant's January 2026 report (opens in new tab) shows what happens when that layer is missing. Attackers called employees pretending to be IT, said the company was updating its multi-factor authentication (MFA) settings, and sent them to a fake login page branded like their company's. The page captured each employee's single sign-on (SSO) password and MFA code, and the attackers then registered their own device for MFA.
Google's Threat Intelligence Group (opens in new tab) described a related pattern in Salesforce in September 2025. During a phone call, the attacker walked an employee through approving a malicious connected app on the real Salesforce site. No Salesforce vulnerability was involved. Google's hardening guidance (opens in new tab) starts from the idea that most employees should not be able to approve an app like that at all: block unapproved connected apps by default, and limit API access to the few users who need it.
Neither attack started with an email, so no email filter had a chance to catch it. In both, the person was one point of failure, and an identity control was another.
The UK's National Cyber Security Centre (NCSC) has been stating the uncomfortable truth since at least 2022: people will sometimes click bad links because clicking links is part of their job, and "being able to spot a phish is not their job (opens in new tab)." Its guidance (opens in new tab) tells organizations to measure how many people reported a phishing message, not only how many clicked it, and not to reprimand people who struggle to recognize one. In the NCSC's words, "Users who fear reprisals will not report mistakes promptly, if at all."
ReliaQuest described the same idea in August 2026, after an attack on its own staff. A caller posing as one of its security employees convinced a staff member to enter their password and approve a push notification. ReliaQuest wrote: "Our defense in depth starts from the assumption that a threat actor will eventually phish someone's account. Phishing works. (opens in new tab)"
The attacker got a brief, view-only session on ReliaQuest's identity dashboard. Every attempt to reach its applications was blocked because device trust only lets ReliaQuest-owned devices connect. The human layer was bypassed, and the controls behind them held.
No single layer, human or technical, stops every attack.
What we want measured
Boards want a click rate, and we understand why. It is one number, and it fits on a slide. The research points to a different number.
Researchers at ETH Zurich (opens in new tab) ran a 15-month study with more than 14,000 employees at one company. About 10%of reports on a simulated phish arrived within five minutes, and 30 to 40% within half an hour. They estimated that in a company of 1,000 people where 100 receive the same phishing email, at least one report would very likely arrive within five minutes. Employees kept reporting for the full 15 months, and automated triage kept the workload small.
The NCSC's own worked example (opens in new tab) shows the same thing. Of 50 phishing emails that got past the filters, 25 were reported, some after someone had already clicked. Those reports were the first sign the attack had gotten through. A 2025 Purdue study (opens in new tab) of 12,511 employees found that in about half of the campaigns it could measure, the first report arrived before the first click.
A click tells you one person was fooled. A report can protect everyone else who got the same message.
Verizon's 2025 Data Breach Investigations Report (opens in new tab) found that people trained within the prior 30 days reported simulated phishing emails at about 21%, against a base rate of 5%, which Verizon calls "a four times relative increase." The effect on clicking was "only 5% relative impact." The Purdue study, which assigned training at random, found no significant training effect on either number. The case for measuring reports does not depend on training. It depends on what a report lets the security team do.
Whatever the channel, the behavior we want is the same: recognize the attack, verify the request through a channel you already trust, and report it. On a text, verifying means calling the person back on a number you already have, not replying in the thread. Reporting means using the path the company provides, so the security team, incident response, or the security operations center (SOC) can act while it is still a report and not yet an incident.
A click rate does not measure any of that. A click records the moment before the other layers of defense have had a chance to work, and on its own it gives the security team nothing to act on. A report is the signal that lets them respond.
All of this research comes from email, where a report button sits in the inbox. On a personal phone, there usually is not one.
Why text messages are harder to defend
A text from someone posing as an executive, IT, or a developer often lands on a phone the company does not manage. There is no email security gateway filtering it, and often no obvious way to report it. The employee has to know a reporting process exists and decide to use it, starting from their lock screen. The same is true on WhatsApp, Telegram, and iMessage. The app changes, but the phone still belongs to the employee.
Text messages with links are a documented and common way to attack employees. CISA's advisory on Scattered Spider (opens in new tab) describes the group posing as company IT staff by phone or text to collect credentials. Group-IB's 2022 report (opens in new tab) on the same activity is specific: employees received text messages with links to pages that imitated their company's Okta login. That threat is real and should be tested. It is also the version of text-message phishing that a lot of testing programs already know how to imitate.
The conversational version works more like a vishing call, and it matches the FBI pattern described above: a short, believable exchange with someone who seems to belong. The ask comes after the target has already answered.
The filters nobody in the company controls
A text message is filtered twice before a person sees it, once by the carrier and once by the phone, and the company controls neither.
Carriers are the first filter. The FCC adopted a rule in 2023 (opens in new tab), in effect since September 2024 (opens in new tab), that requires mobile carriers to block, "at the network level," texts that claim to come from invalid, unallocated, or unused numbers. The order explains why texts get this attention: "consumers read nearly all texts, and do so nearly immediately." The wireless industry's own messaging guidelines (opens in new tab) add that providers "may filter or block Unwanted Messages before they reach Consumers." The FCC rule only covers SMS and MMS, though, "not over-the-top (OTT) messaging, such as iMessage and WhatsApp or Rich Communications Services (RCS)." That is where many of these conversations end up.
The phone is the second filter. For years, Apple has said (opens in new tab) that if you turn on its filter for unknown senders, you can't open their links until you reply or add them as a contact. Apple confirmed the reply behavior to reporters in January 2025 (opens in new tab), and by then some smishing texts were asking people to reply "Y." Both major platforms now go further with suspected spam. In iOS 26, spam filtering is on by default (opens in new tab), and Apple says "you can't reply to spam messages or click on links in a spam message until you move it to your Inbox." Google Messages warns on links in suspected spam (opens in new tab) and stops the person from visiting the site unless they mark the message as not spam. Apple also notes that its message filtering "doesn’t apply to any sender you’ve replied to three times or more (opens in new tab)." A conversational attack gets there on purpose.
That means a click rate on a text message reflects the carrier’s filtering and the phone’s settings as much as the person. Links are still worth testing, but a click is only one way an attack starts.
The public reporting shows the same pattern across channels: a believable story, trust built over time, and a request that comes later. The medium will keep changing, but the influence underneath will not, which is why we test the conversation and not only the link. And because employees are one layer of defense, we measure whether people recognize, verify, and report what they see, so the layers behind them can respond.
Recognize it. Verify it. Report it.
The human is one layer. Not the only one.
To see how Doppel tests conversational smishing and measures reporting, not just clicks, click here to request a demo.