Wait... What the Fuck is the Fraud Triangle?

If you have spent any time trying to parse the breathless announcements coming out of frontier artificial intelligence labs lately, you are likely suffering from acute cognitive exhaustion.

You are told that an unreleased model—locked inside a cryptographic vault, codenamed like a Cold War stealth bomber, and briefed to trembling cabinet ministers behind closed doors—is so profoundly powerful that it threatens the biological and digital foundations of civilization. To prove this, the lab releases a 240-page technical system card filled with synthetic evals, differential equations, and portentous warnings about catastrophic thresholds.

If you attempt to engage with this on the lab's terms, you lose. You get dragged into a swamp of speculative metaphysics: What is machine consciousness? Can a transformer exhibit autonomous agency? How do we calculate the probability of extinction?

Stop. Put down the philosophy papers. Step away from the neural network textbooks.

The tool you need to understand what is happening wasn't written by a computer scientist at Stanford or an alignment theorist in Berkeley. It was formulated in 1953 by an American criminologist named Donald R. Cressey.

It's called the Fraud Triangle. And once you look through it, the entire frontier AI circus stops looking like the dawn of a new species and starts looking like an ordinary, high-leverage corporate shakedown.

The 1953 Epiphany

Donald Cressey was a sociologist who wanted to know why seemingly respectable people violate financial trust. He didn't study criminal masterminds or desperate street thieves; he studied embezzlers, corporate accountants, and executives who had clean records, solid reputations, and high institutional standing.

Cressey discovered that occupational fraud rarely happens because a villain wakes up and decides to do evil. It happens when three distinct structural conditions collide simultaneously:

1. Perceived Pressure: A non-shareable financial problem or crisis that the individual or institution cannot resolve through normal, open market mechanisms without facing ruin.
2. Perceived Opportunity: A flaw in oversight, an absence of controls, or an epistemic gap that allows the perpetrator to manipulate the ledger without getting caught.
3. Rationalization: An internal ethical framework that allows the perpetrator to justify the deception to themselves, recasting a self-serving violation as an acceptable—or even noble—necessity.

For seventy years, forensic accountants, auditors, and regulators have used Cressey's triad to autopsy everything from collapsed savings-and-loans to Enron and Wirecard.

Here is why it applies with surgical precision to the current frontier AI landscape.

Leg 1: Pressure (The Commoditization Abyss)

Frontier AI laboratories are trapped in an economic horror movie.

To train and serve each generation of foundational models, labs like Anthropic burn through billions of dollars in capital provided by hyperscale conglomerates and venture syndicates. That money is not an unconditional grant for the advancement of human knowledge; it is an investment predicated on the extraction of monopolistic rents. The investors expect these labs to become the indispensable tollbooths of the modern economy.

There is only one problem: raw inference is in freefall toward commoditization.

Language models, at their core, produce tokens. When an open-weight model—runnable on local consumer or enterprise silicon without recurring API bills, corporate surveillance, or arbitrary platform neutering—approaches closed commercial endpoints on practical coding and writing tasks, the economic floor drops out. The customer doesn't need to replicate the multi-billion-dollar frontier; they just need to stop paying the toll.

A lab cannot service billions in compute debt by selling commodity autocomplete. If an enterprise board discovers that a local, fine-tuned model handles 95% of its workflow for the cost of electricity, Anthropic's multi-billion-dollar valuation evaporates overnight.

The pressure is immediate, structural, and non-shareable: the business must manufacture an artificial moat or face terminal insolvency. It cannot survive by merely selling a better utility; it must sell something that open-source software is legally prohibited from becoming. It must sell existential liability insurance.

Leg 2: Opportunity (The Gated Epistemic Castle)

To escape the commoditization trap, the lab must convince governments, procurement officers, and enterprise boards that un-governed, open compute is an unacceptable catastrophic hazard.

This requires demonstrating that the lab's newest models possess world-threatening capabilities—such as autonomous cyberwarfare or biological weapon synthesis—that demand state licensing, mandatory compliance frameworks, and strict containment.

This is where Cressey's second leg emerges: Opportunity. In classic occupational fraud, opportunity means the perpetrator holds the keys to the lockbox and grades their own ledger. In frontier AI, opportunity is achieved by locking out the scientific method under the banner of "containment."

• The Unfalsifiable Hazard: The lab declares that its model represents an AI Safety Level 3 (ASL-3) threat. But when independent academic researchers or security analysts ask to inspect the weights, test the prompts, or examine the exploit payloads to verify whether the claim is true, the lab slams the door: "We cannot show you the evidence, because the evidence itself is a weapon." The withholding of the model is marketed as proof of its omnipotence, while that supposed omnipotence is cited to justify withholding the model.
• The Captive Cartel: To manufacture an audit trail, the lab invites selected partners into a closed room under strict NDAs (the playbook seen in Project Glasswing). The guest list tells the whole story: Amazon (protecting its multi-billion-dollar stake), Nvidia (protecting its silicon monopoly), Wall Street banks (buying frontier-hardened bragging rights), and the UK AI Security Institute (a captive state bureaucracy whose budget depends on the threat being catastrophic).
• Access as Epistemic Remuneration: The evaluators inside that room are not independent auditors with subpoena power; they are handpicked guests inducted into a mystery: "I see the light! Anthropic has chosen me!" Selection confers scarce prestige, early access, and regulatory relevance. The evaluator acquires a direct professional stake in the authority of the institution that selected them. They will argue vigorously about sandbox parameters, but they are institutionally incapable of asking the fatal question: is the sandbox itself a commercial stage prop?
• The Empirical Shell Game: Behind the terrifying press releases lies standard, narrow automated fuzzing. In Anthropic's own Firefox JavaScript-shell evaluation, Mythos claimed 72.4% full code execution; remove just two dominant bugs, and that success rate plummets to 4.4%. Out of 23,019 headline-grabbing findings, only 75 high/critical bugs were actually patched—a mere 0.33% of the pile. The public gets the terrifying denominator; the footnotes get the negligible reality.

The lab manufactures the threat metric, controls the testing harness, withholds the data from public falsification, and then presents its own systems card to legislative committees as objective proof that it must be granted a statutory monopoly.

Leg 3: Rationalization (The Longtermist Priesthood)

The third leg of Cressey's triangle is the most dangerous, because it explains why the executives and researchers involved can look you dead in the eye without flinching.

They do not believe they are running a hustle. They believe they are saving the world.

The frontier AI ecosystem is heavily saturated with the ideology of Effective Altruism, longtermism, and existential risk soteriology. Within this worldview, artificial general intelligence represents a threshold event: unaligned compute will lead directly to the extinction of the human species, while "properly aligned" compute will shepherd humanity into a post-scarcity cosmic paradise.

From that foundational axiom, the moral calculus is automatic:

• If an open-source developer publishing uncensored weights risks building an "existential bio-hazard," then lobbying Congress to criminalize independent model training isn't anti-competitive cartelization—it is species disarmament.
• If an unreleased model's automated bug-fuzzing capabilities have to be exaggerated to terrorize lawmakers into passing licensing moats, that isn't deceptive marketing—it is a precautionary duty.
• If enterprise clients must be scared into paying a massive safety premium to fund Anthropic's next compute cluster, that isn't extraction—it is funding the righteous side of the AGI race.

When you genuinely believe that the survival of eight billion human souls depends on your specific company maintaining market dominance over your "reckless" rivals, standard commercial ethics become quaint irrelevancies. The noble end rationalizes every administrative deception.

The Audit We Actually Need

The tragedy of the current moment is not that frontier laboratories are executing this playbook. Corporations under terminal capital pressure have sought regulatory capture and manufactured artificial moats for centuries.

The tragedy is that the people whose public duty it is to scrutinize them—parliamentary committees, antitrust regulators, national consumer watchdogs, and technology journalists—have completely vacated their posts. They have allowed themselves to be blinded by the vocabulary of high-energy physics and moral philosophy. They treat a corporate PR rollout as if it were a dispatch from the oracle of Delphi.

You do not need to understand backpropagation or transformer self-attention to see through this. You only need to know how white-collar institutions behave when the cash is running out, the ledger is hidden from view, and the leadership believes they are on a mission from God.

The Fraud Triangle tells us exactly what to do:

1. Ignore the theology: Stop debating whether the autocomplete engine has a soul or will build a synthetic pathogen in 2035.
2. Break the information monopoly: Strip away the legal shields. If a company claims a model is too dangerous to release, that claim cannot be cited as grounds for public regulation or commercial exclusivity until the underlying data is made available for unrestricted, adversarial testing by un-conflicted third parties outside the NDA tent.
3. Follow the money: Treat risk announcements not as scientific discoveries, but as representations made to extract investment, defend equity valuations, and secure legislative protection against cheaper competitors.

The emperor is dressed like shit, thinks like a psychopath, and has a playbook slightly younger than my mum.