The event was real, the registration page was not: operator-driven relay phishing against Taiwan’s research community
Cisco Talos has published a spear-phishing campaign against people affiliated with Taiwan-based research organisations. The relevant finding is architectural: a persistent operator channel decides which multi-factor screen the victim sees, which turns the choice of second factor into a branch in the attacker's state machine.

Cisco Talos published research on 8 October 2026, authored by Joey Chen, on a spear-phishing campaign it tracks as UAT-11985 and observed in mid-2026. The targets were individuals affiliated with research organisations based in Taiwan [1]. The lure is competent social engineering. What matters outside Taiwan is the machinery behind it.
The lure
The emails invited recipients to public events, using real event titles, dates and venues as cover, and impersonated the Taiwan European Union Centre, the NCCU Institute of International Relations and the Taiwan Research Institute. When recipients contacted those organisations, none could confirm the three senders worked for them, and Talos concludes the sender identities were probably fabricated [1].
Each email has three parts: an elaborate geopolitical framing, flattery aimed at the recipient, and event logistics. The registration link displays what looks like a legitimate Google Forms address while the underlying link points to a phishing site on a third-party platform [1]. The operators also distributed modified event posters carrying malicious QR codes, which can reach people who never received the email [1]. No victim, email or poster counts are published [12].
On AI authorship Talos is careful. It says it cannot conclusively determine whether a large language model wrote the emails, but reports strong evidence of AI-assisted production: near-identical structures across different topics, formulaic policy jargon, interchangeable praise, and the likely reuse of a prompt template [1]. That is a statement about production economics, not capability. The marginal cost of a credible, individually framed invitation has fallen, and the limit on how many a campaign can send is no longer the writing.
The operator is in the loop
The kit replicates the Google sign-in page and relays traffic to the real authentication service. It runs two channels. Transient data — browser fingerprints, identifiers, credentials, heartbeats — goes out over ordinary web requests. Alongside that, a persistent socket stays open to the operator, who pushes state updates deciding which multi-factor screen the victim sees next. The server checks whether a passkey flow is enabled on the account and relays the challenge type accordingly [1].
That is the finding. In a static credential-harvesting page, multi-factor authentication is an obstacle the attacker must anticipate in advance. In a relay with a live operator it is a branch: the kit learns which factor the real service is asking for and presents the matching screen. Talos states the end goal as complete authenticated session tokens, which permit account takeover [1]. The password is an intermediate artefact.
The architecture is not confined to espionage. Talos research of 13 August 2026 describes an undocumented phishing framework it calls JWR, which impersonates checkout and login flows for payment platforms and which Talos assesses with medium confidence to be a variant of an existing phishing-as-a-service platform. There the operator can issue more than forty instructions, and typed input is streamed as entered, so partial card numbers and codes are visible before submission. Observed lures impersonated road-pricing authorities, postal services and courier brands [2]. Two operator populations, one design.
What this does to the second factor
The technique is catalogued. MITRE ATT&CK describes adversary-in-the-middle as positioning between networked parties to sniff traffic, manipulate data or replay credentials, with stealing access tokens and session cookies among the goals [5]. The companion technique on stealing web session cookies states that session cookies may bypass some multi-factor authentication, names relay phishing proxies among the acquisition methods, and lists hardware tokens bound to the login domain as mitigation [6]. The display-versus-destination mismatch used here is the spearphishing-link sub-technique [7].
Microsoft’s figures for the first quarter of 2026 give the volume context. It detected about 8.3 billion email-based phishing threats in the quarter, with monthly volume easing from 2.9 billion in January to 2.6 billion in March. Credential phishing was the goal of 89 per cent of malicious payloads in January, 95 per cent in February and 94 per cent in March, while traditional malware delivery fell to 5 or 6 per cent. QR code phishing grew from 7.6 million attacks in January to 18.7 million in March, a 146 per cent rise that Microsoft calls the fastest-growing vector [3]. The mailbox is now almost always the objective rather than a delivery route to something else, and the poster vector is not an eccentricity.
Google’s published position, from July 2025, is that phishing and credential theft drive 37 per cent of successful intrusions and that cookie and token theft has risen sharply. Its two answers are passkeys, which are phishing-resistant and scoped to a single site, and device-bound session credentials, which tie a session cookie to the device that signed in. At publication passkeys were generally available across its enterprise estate, while device binding was in open beta on one browser and operating system [4]. Passkeys address the credential half of a relay and device binding the stolen-token half; neither is the default, and an account holding one without the other stays exposed on the side it lacks.
Attribution, held where the researchers held it
Talos assesses with moderate confidence that the kit’s interface was first written in Simplified Chinese and later adapted for Traditional Chinese and English, which it reads as indicating a developer whose primary working language is Simplified Chinese. It names no state sponsor, and nothing further should be inferred from the language assessment [1][12].
The targeting pattern is not new, and the figures diverge in ways worth publishing. Taiwan’s National Security Bureau reported an average of 2.63 million intrusion attempts a day against critical infrastructure in 2025, 6 per cent above 2024’s 2.46 million, itself up from 1.23 million in 2023, attributing 18 per cent of attacks to social engineering and 57 per cent to exploitation of software and hardware flaws [8]. A separate figure of 2.8 million intrusions a day during 2025, up 17 per cent, was reported from local coverage of the Bureau’s work on government networks rather than critical infrastructure [10]. The two measure different populations and both are given here. Talos separately identified a China-nexus intrusion set in March 2026 while investigating spear-phishing against Taiwan’s academic, think tank and civil society policy community [11], and in July 2026 Taiwan’s Ministry of Digital Affairs confirmed an intrusion of government agencies involving at least 85 user accounts, which the firm that detected it did not attribute [9].
What to do
- Verify an invitation against the organisation’s own published contact details, not by replying to the email.
- Treat a QR code on a printed poster as an untrusted link, because it carries none of the sender context an email does.
- Enrol passkeys or origin-bound hardware security keys for anyone whose mailbox has policy, research or commercial value, and make them the only accepted factor for those accounts.
- Alert on session-token use from a new address, device or user agent without reauthentication, which is the signal a relay leaves after the sign-in itself looks normal.
- Shorten session lifetimes on accounts that matter and audit third-party application grants, since a relayed session is often used for persistence that survives a password reset.
- For published coverage the research names a ClamAV signature, Html.Phishing.UAT11985-10060614-0, and Snort identifiers 1:67198 and 7:31 [1].
The operational lesson concerns what an organisation publishes. A conference programme, a speaker list, a venue and a date are exactly the material that makes a fabricated invitation credible, and they are published on purpose. The answer is not to publish less. It is to assume anything on the public programme will be reused as a lure, to give the people named on it credentials that cannot be relayed, and to tell the audience in advance how a genuine registration link will look. This campaign spends its sophistication on the authentication step and almost none on delivery, because that is where the defence sits.
Sources
Every R3KONX article cites its primary material. 12 sources, in order of first citation. Links open the original publication.
- UAT-11985: AI-assisted event lures delivering real-time Google AitM phishing Cisco Talos · 2026-10-08
- Dissecting the JWR phishing framework Cisco Talos · 2026-08-13
- Email threat landscape: Q1 2026 trends and insights Microsoft Security Blog · 2026-04-30
- Defending against account takeovers: top threats, passkeys and DBSC Google Workspace Blog · 2025-07-30
- T1557: Adversary-in-the-Middle MITRE ATT&CK · 2026-10-09
- T1539: Steal Web Session Cookie MITRE ATT&CK · 2026-10-09
- T1566.002: Spearphishing Link MITRE ATT&CK · 2026-10-09
- Chinese cyberattacks rising: NSB report Taipei Times · 2026-01-05
- Taiwan targeted in AI-driven hacking campaign Taipei Times · 2026-08-14
- Taiwan reports surge in Chinese cyber activity and disinformation efforts The Record · 2025-10-14
- Researchers find Chinese hacking campaigns targeting AI firms, Asian governments The Record · 2026-10-01
- UAT-11985: Real-Time Google AitM Phishing Explained NetmanageIT CTO Corner · 2026-10-08
Researched and written by the R3KONX analysis desk from the cited primary material. Three caveats apply. The research gives no victim, email or poster counts, and none are estimated here. The assessment of the developer's working language and the assessment that the lures were produced with AI assistance are the researchers' own and are reported as such. The defensive measures in the final section are this desk's, not the researchers', except where the detection identifiers are quoted. Corrections to event@r3konx.asia.
