# The Oldest Attack Vector in the Room Has No Patch
She walked onto the DEF CON 34 stage in a deerstalker cap, plaid shawl, and pipe — and made a room full of hackers feel like they were being out-thought by a Victorian detective who died on paper in 1893.
That was the point.
Elizabeth Rasnick, an assistant professor at the University of West Florida's Center for Cybersecurity and AI, spent her session arguing that Sherlock Holmes wasn't just a crime solver. He was the original social engineer. The costume wasn't cosplay — it was the demonstration itself. Rasnick had transformed into Holmes well enough to briefly blur the line. And social engineering is exactly that: blurring lines until the target can't tell what's real.
## A Playbook That Predates the Internet by a Century
Holmes's methods, as Conan Doyle described them, map almost surgically onto what CISA would call pre-attack reconnaissance and manipulation today. He studied targets before engaging. He built cover identities — including, at one point in the original stories, maintaining a fake engagement to a housemaid for weeks to extract information. He created urgency, exploited curiosity, and adapted when his initial approach failed.
That's not metaphor. That's a kill chain.
Rasnick laid out the parallel explicitly: threat actors today operate on the same six-stage logic — know the target, become believable, create a reason to act, exploit emotion, observe behavior, adapt. The tools have changed. The psychology hasn't moved in 140 years.
Modern attackers use OSINT tooling to map out a target's social graph before a spear-phishing email ever gets drafted. Holmes did the same with legwork and his "Irregulars" — a network of street informants who could go unnoticed anywhere in London. He read behavioral tells from physical observation; today's threat actors do it by scraping LinkedIn, correlating leaked databases, and monitoring public Slack channels. Scale looks different. The instinct is identical.
## Why AI Doesn't Change the Root Problem
The framing that AI "supercharges" social engineering is technically true and contextually insufficient. Yes, language models let attackers draft phishing emails without grammatical tells, clone voices in minutes, and synthesize deepfake video convincing enough to fool a finance team into wiring money to the wrong account. The 2024 Arup deepfake incident — where a Hong Kong employee transferred $25 million after a video call with a fake CFO — showed exactly where this goes at scale.
But the AI angle, covered endlessly, obscures the more durable lesson from Rasnick's talk: predictable behavior is what makes social engineering possible. Fear, urgency, curiosity, authority — these aren't bugs in human cognition that AI exploits. They're features that attackers have been weaponizing since before computers existed.
The question for defenders isn't "how do we detect deepfakes?" — though that's urgent too. It's "why does urgency from an authority figure still override verification protocols in organizations that have run phishing training for a decade?"
Rasnick named it plainly: trust is the real attack surface.
## The Training Problem No One Wants to Admit
Security awareness training has been the industry's answer to the human element for twenty years. The results are mixed at best, and the honest answer is that we don't have strong evidence it moves the needle durably.
What training typically does is teach employees to recognize the last attack — the exact phishing email style from three years ago, the obvious Nigerian prince energy, the easy-to-spot urgency. It doesn't prepare them for a call from their CEO's voice, synthesized from six months of earnings calls and investor podcasts. It doesn't prepare them for a slow, multi-week relationship-building operation that never asks for anything unusual until the one moment it matters.
Holmes would have recognized the gap immediately. He didn't study criminals to recognize their past methods. He studied people — their routines, their vanities, their predictable responses under pressure — to anticipate what would work on them specifically.
The industry's answer to social engineering keeps being more awareness training. Holmes would have asked who's profiling the attacker's methodology and what controls exist that don't depend on the target being alert at the exact wrong moment.
## What Defenders Actually Get From This
The practical takeaway from Rasnick's framing isn't "hire Sherlocks." It's that the behavioral constants Holmes exploited — and that attackers exploit — are stable enough to design against.
A few concrete angles:
Verification architectures, not just verification reminders. If urgency and authority are the attack vectors, then high-stakes actions (wire transfers, credential resets, access grants) need out-of-band confirmation built into the process. Not a reminder to "confirm unusual requests" — an actual friction point that makes bypassing confirmation structurally difficult.
OSINT audits of your own organization. Before attackers build their target dossier from your company's public footprint, you should. LinkedIn org charts, conference speaker bios, job postings, GitHub commit histories — these collectively hand an adversary a ready-made social graph. Most organizations have no idea what their public profile looks like assembled.
Red team with pretext, not just technical exploits. A penetration test that doesn't include social engineering is measuring half the attack surface. Holmes would have walked through the front door.
The session at DEF CON hit at something the industry keeps circling without landing on: technical controls are necessary and insufficient, and the human layer isn't a problem to be solved with annual training. It's an attack surface that requires the same architecture of defense-in-depth applied to everything else.
The attacker's toolkit evolves. The underlying exploit — human predictability under pressure — doesn't. That asymmetry is the actual threat model.
---
## HackWire Analysis
Rasnick's DEF CON session lands at a specific moment worth noting: AI-driven social engineering is no longer a theoretical threat, and the industry is still largely responding with the same playbook it had before the capability shift.
The deeper pattern here is one the security community hasn't fully absorbed. Social engineering has historically been treated as a "soft" problem — addressable through training, policy, and culture — while technical teams focused on the hard problems of network segmentation, patch management, and detection engineering. That division made some sense when the cost of a polished, convincing impersonation was high. A believable deepfake video of an executive required production resources. A natural-sounding voice clone required custom tooling.
Neither of those is true anymore. The cost curve collapsed. What once required a nation-state-level operation now requires a free trial account and an afternoon.
That changes the severity calculus for social engineering in ways most risk frameworks haven't caught up with. The incidents we're seeing — the Arup case, MGM's vishing-triggered ransomware incident in 2023, dozens of BEC campaigns running against mid-market companies right now — aren't aberrations. They're the early signal of a much larger wave.
Rasnick's Holmes framing is valuable precisely because it cuts through the AI hype to the durable truth: the attack works because humans are predictable, not because the technology is clever. Defenders who focus exclusively on detecting the tool miss the point. The tool changes. The exploit doesn't.
The organizations that will hold up against next-generation social engineering aren't the ones with the best phishing simulations. They're the ones that have built verification into process rather than relying on vigilance — and that have mapped their own social graph as carefully as an attacker would.
— HackWire Editorial
---
## Related Coverage