The Man Who Wasn't There
A security researcher who did everything right is erased by a counterfeit of himself — his face, his voice, his fingerprints in the code — until the world agrees, with footage, that the real one was the criminal.
01 · ANATOMY OF THE FRAMENobody hacked Elias Marchetti
That is the part people get wrong about this kind of attack, and the reason the episode opens where it does. There is no breach in this story. No stolen password, no zero-day, no one crouched over a terminal peeling a firewall apart. Everything used against Elias was already public, already legal to look at, and most of it he published himself.
What follows is the same map a defender would draw after the fact — the techniques named, in the order the episode uses them. The IDs are MITRE's. The reading is ours.
It starts with reading TA0043 · Reconnaissance
Before anyone generates a single frame of fake video, somebody reads. Conference talks. Commit history. The company blog. A podcast appearance where he laughs at his own joke for four seconds. Job postings that quietly describe the org chart. This is T1589 and T1589.001 — gathering identity information and personal identifiers — sitting alongside T1591, gathering victim organisation information: who signs off, who gets believed, who is too busy to check.
None of those facts is a secret. The power is in correlation — thousands of small true things fused into inferences the target never published. And here is the honest part the show refuses to hype: through 2024–25, AI made this faster and broader, not smarter. The link-analysis tooling that does the heavy lifting predates the current wave by years. What changed is the cost of doing it to everybody, all the time, instead of to the handful of people somebody was already paid to care about.
Then they build a man T1656 · Impersonation
The material he published becomes the training set for a counterfeit of him. T1656 is the technique's dry name; the episode's name for it is the thing in the car. In MITRE's AI-specific matrix this is closest to AML.T0043, crafting adversarial data — generating the media itself through an ML-enabled service.
Say the fit out loud, because the mismatch matters: ATLAS catalogues attacks against machine-learning systems. This is misuse of generative AI as a tool. Nobody in this story attacks a model. They rent one. That is why the enterprise matrix carries this episode and ATLAS only brushes it — AML.T0048, External Harms, with its reputational and societal sub-entries, is the nearest abstraction and it is still a loose one.
A lie needs a history T1585 · Establish Accounts
A counterfeit that appears from nowhere gets checked. A counterfeit with two years of ordinary posting behind it does not. T1585 and T1585.001 cover standing up that footprint — the digital dust that makes the fake survive the ten seconds of scrutiny anyone actually gives it. This is the least cinematic step and the one that does the most work.
Delivery is the weapon T1598 · T1566.002
A perfect fake sitting on a hard drive harms nobody. The frame becomes an attack when it is delivered — T1598, phishing for information, and T1566.002, spearphishing through a service the target's colleagues already trust. It arrives inside the channels that carry real news about real people. Then, where impersonation converts into actual access, you get T1078, valid accounts, and the reputational attack becomes an operational one.
02 · WHY NOBODY CAUGHT ITThe defences you would expect to fire
Ask most people how you beat a deepfake and they will describe a detector: a box you feed the video into that lights up red. That box is not what saves you, and the episode is built on the gap.
What forensics can honestly say
Media forensics is real and it works — on physiological and signal tells. Breathing cadence that is absent or impossible. Room tone that changes mid-sentence. Editing seams. In video: blink and gaze that never quite land, lighting and reflections that disagree with the room, remote pulse estimation that finds no pulse.
But note what those findings actually establish. They flag inconsistency with a genuine capture. They do not "prove AI." That distinction is not pedantry — it is the difference between a defensible expert opinion and one that collapses under cross-examination, and any tool that claims the stronger version is overselling.
Provenance, and the trap inside it C2PA · SynthID
Content provenance standards — C2PA / Content Credentials — let a camera cryptographically sign an image at capture. Generator watermarks like SynthID mark output as machine-made. Both are genuinely useful, and both are useful in exactly one direction.
Absence of provenance is not proof of forgery. A signed asset attests itself. An unsigned one tells you nothing — which is every photograph ever taken before the standard existed, and every honest file since that passed through software that stripped the manifest.
Marks can be removed. Nothing compels a hostile generator to apply one in the first place. Treat provenance as a strong positive signal and a worthless negative one, and you will read these systems correctly. Treat it as a lie detector and you will convict innocent people.
The standards that govern the room, not the file
ISO/IEC 30107 is the presentation-attack-detection family — the liveness checks behind identity verification. It becomes central from Episode 4 on, when the season turns to synthetic identity kits. Here it is the door nobody knocks on. NIST's AI Risk Management Framework supplies the vocabulary for the institutional failure in the back half of this episode: not a technical failure at all, but nobody being accountable for the decision to believe a machine. And the OWASP Top 10 for LLM Applications — LLM01, prompt injection, chief among them — is not this episode's centre of gravity. It becomes one from Episode 2. Listed here so the season's map is honest about where it starts.
03 · WHAT ACTUALLY BREAKS A FRAMEThe grubby, unglamorous answer
In the real cases this episode is built from, frames do not collapse because someone proved the media synthetic. They collapse because investigators traced who sent it, and from where.
Email headers. IP addresses. Device fingerprints. A recovery phone number attached to an account somebody set up in a hurry. The attacker's operational-security failure is what undoes the attack — not the quality of the fake, and not a detector. It is the least satisfying answer and the most reliably true one, which is why it is the beat the episode gives its detective work to.
04 · THE CRUELLEST TRUE DETAILThe liar's dividend
Here is the finding that should worry you more than any technique above.
The more widely people understand that deepfakes exist, the less a framed man's truthful denial is believed.
Every public warning about synthetic media — including, in its small way, this one — deposits a little more ambient doubt into the world. That doubt is what let the fake pass in the first place. It is also what discounts Elias when he says that wasn't me. The same fog does both jobs. Researchers call it the liar's dividend, and there is no clean countermeasure, because the countermeasure and the disease are the same substance.
And the scarier, more accurate horror is not the flawless counterfeit. In life the fake is usually imperfect — and it works anyway, because institutions act on it before anyone checks. Speed and outrage do the damage. Forensics finds the seams later, to an audience that has already moved on.
05 · WHERE THE FICTION LEANSWhat we made up, said plainly
This show keeps a research brief with a verification ledger for every technique it dramatises, so it can answer this question instead of dodging it.
Real and usable as written: a specific, named person can be convincingly framed by a deepfake of their own voice. The harm — suspension, threats, the presumption of guilt — lands well before the truth catches up. The honest denial gets discounted. All three are documented.
The novel's move: getting a living person declared dead on synthetic evidence is not a documented event. The episode fuses two real dynamics that reality keeps in separate lanes — deepfake framing, and wrongful "civil death" through records error. That fusion is dramatisation, and if you ask whether it has happened, the honest answer is that the pieces have; the fusion is ours.
ATT&CK is revised continuously and ATLAS more so. Every mapping above is taken from this show's own technique briefs — briefs 01 (OSINT / reconnaissance) and 02 (synthetic-identity framing) — each carrying a per-claim verification ledger. Where a widely-cited source article had a mapping wrong, the brief records the correction rather than repeating it. Re-verify against the live matrices before you rely on any ID here in your own work. This page is accurate about what a technique does and what it costs. It is not, and will never be, a how-to.