AI Spearphishing: How OSINT and Threads Build Lures

AI spearphishing is phishing written by a language model for one person or a small group, built from public data about that person's role, manager, vendors and active projects. Attackers use this context to write a single, credible message rather than mass-blasting a generic template, which is why old red flags like typos no longer apply.

Key takeaways

What is AI spearphishing, exactly?

AI spearphishing differs from bulk phishing in one key way: the message is written for a specific person using specific facts about them. According to AegisAI's State of the AI Attack report, the volume of this category grew 88% and its share of all confirmed phishing climbed from 9.7% to 12.3%.

The report describes the technique plainly: attackers gather "their role, their manager, their vendors, a deal or project in progress" and then use a language model to generate a message built from that real context. The output is not a template with a name swapped in. It is a message that reads like it was written by someone who already knows the target's world.

This matters because detection systems and employees have spent years learning to spot generic phishing. AI spearphishing is built specifically to not look generic.

How do attackers gather OSINT to build a lure?

Open-source intelligence (OSINT) is the raw material. Threat actors construct persuasive pretexts by harvesting public intelligence, including job titles, reporting lines, professional networking updates, corporate press releases and company news. Every public data point becomes a variable that can be weaponized in a message.

AegisAI's own red-teaming work shows how fast this gathering happens. In a breakdown of how an AI attack is built, an AI agent was asked to research a real person and returned a structured dossier in seconds, including a recent funding announcement, a podcast appearance and an upcoming conference speaking slot. None of this required a skilled human researcher. The agent identified which facts made good pretext material on its own.

This is the shift security teams need to internalize: the OSINT step that once took a human attacker hours of LinkedIn and news searching now takes a model seconds, and it can be repeated across an entire employee directory.

Why is thread hijacking the top AI spearphishing lure?

According to AegisAI's Q3 2026 report, thread hijacking is the top AI spearphishing lure, accounting for 20.3% of volume. Thread hijacking means inserting a malicious request into a real, ongoing email conversation, either by compromising an account in that thread or spoofing a participant convincingly enough to blend in.

An AI model makes this easier to scale because it can read the existing thread, match its tone and terminology, and produce a reply that fits seamlessly. The attacker does not need deep familiarity with the business relationship. The model infers it from the conversation history.

How does vendor impersonation fit into AI business email compromise?

Vendor impersonation is one of the most damaging applications of this technique because it targets payment workflows directly. A convincing vendor-change request, delivered inside a thread that already discusses an invoice or a deal in progress, does not need to trick the recipient into visiting a fake login page. It only needs the recipient to update a bank account number or approve a wire.

This is a core driver of AI business email compromise losses. Industry-wide figures collected in StationX's phishing statistics cite FBI IC3 data showing $2.77 billion in reported BEC losses from 21,442 complaints in 2024, and an average BEC attack cost of $4.67 million (IBM, 2025). These losses are not driven by crude, obviously fake emails. They are driven by requests that look exactly like the vendor correspondence an employee already expects.

AegisAI's guidance on preventing business email compromise and dedicated BEC attack prevention approach both treat payment and vendor-change requests as high-risk actions that require verification, regardless of how legitimate the surrounding thread appears.

Why don't old phishing heuristics work anymore?

Security awareness training has taught employees to look for specific signals: poor grammar, generic greetings, mismatched sender names, urgent and oddly worded requests. AI spearphishing is built to defeat exactly these checks.

According to AegisAI's report, AI-generated phishing now reads more personalized than human-written phishing. A December 2024 FBI IC3 advisory explicitly warned that criminals are using generative AI to produce content free of the grammatical errors and foreign-authorship signals that once helped recipients identify fraud.

This trend is also visible at the industry level. Dark Reading reported that phishing overtook exploitation of external vulnerabilities as the top initial access vector in Cisco Talos incident-response engagements during Q1 2026, with more than a third of investigated compromises starting as successful phishing.

How should security teams respond?

Typo-and-greeting heuristics were never going to survive contact with a model that writes fluent, context-aware English by default. The more durable response has three parts:

AegisAI's broader writeup on the new world of AI spearphishing covers how detection needs to move from pattern-matching on language to evaluating the legitimacy of the request and the relationship behind it. For how the same report's numbers play out at machine speed, see why pre-delivery detection wins on AI phishing response time and why DMARC-passing emails can still be phishing.

FAQ

Next step

If you want to see how much of your organization's public footprint an attacker could turn into a working lure, book a demo of AegisAI's AI red-teaming agent and see your exposure before someone else does.