Deepfake Attacks Keep Working Because We Keep Detecting Instead of Proving
Detection is a classification problem, and its error rate grows as the adversary improves. Cryptographic identity proof doesn't care how realistic the fake is — it never relies on what someone looks or sounds like.
In 2025, deepfake-related fraud losses in the US hit $1.1 billion, tripling from $360 million the year before. Deepfake-enabled vishing surged 1,600% in a single quarter. The security industry’s primary response has been better detection — train models to spot synthesized video, analyze audio for artifacts, layer on liveness checks.
The problem: detection is a classification problem, and classification error rates grow as the adversary improves. Deepfake detection tools are 45–50% less accurate in real-world conditions than in the lab. Human accuracy at spotting high-quality deepfake video sits at 24.5%. On episode 92, Jasson Casey, CEO of Beyond Identity, framed the issue structurally: detection is reactive, always responding to the latest generation technique. Cryptographic identity proof is deterministic — it doesn’t care how realistic the fake is, because it never relies on what someone looks or sounds like.
The $25 million video call
In early 2024, a finance worker at British engineering firm Arup received a message appearing to come from the company’s CFO requesting an urgent transfer. The worker was suspicious — correctly. Then they joined a video call. Every participant, including the apparent CFO, was a real-time deepfake. The faces looked right. The voices matched. The worker made 15 wire transfers totaling $25.6 million.
Detection worked at the email stage; the worker spotted the phishing. It failed completely at the video call, because the deepfakes were good enough to override human judgment. This is the logical endpoint of the detection paradigm: as generation quality improves, the window in which detection works shrinks. A voice clone now needs only three seconds of sample audio for an 85% match. The Arup attack didn’t exploit a software vulnerability — it exploited a trust architecture built on visual and auditory recognition. The fix isn’t better detection. It’s a different trust architecture.
Identity is the real security boundary
Casey frames deepfakes as a symptom of a deeper failure: most identity verification ultimately relies on something that can be faked — a face, a voice, a password, a one-time code. “Instead of doing the root cause analysis of the event, we much rather prevent the event from happening in the first place.” His thesis is that identity, not infrastructure, is the real security boundary. If authentication doesn’t depend on recognizing a face or voice, a perfect deepfake of that face or voice is irrelevant to the security decision. This is the shift from detection to proof: detection asks “Is this real?” — which gets harder every year. Proof asks “Can this device produce a valid cryptographic signature?” — which has a mathematically verifiable answer that doesn’t degrade as AI improves. Organizations building zero trust architectures face this choice directly.
How cryptographic proof makes deepfakes irrelevant
Three technologies converge to make “prove, don’t detect” practical. FIDO2 passkeys authenticate through cryptographic key pairs bound to physical devices; the private key lives in a TPM and never transmits. Had the Arup worker been required to authenticate each call participant through a FIDO2 challenge, the deepfake video would have been irrelevant — the attacker would have needed physical possession of each person’s device. This is the same device-bound-credential architecture that eliminates credential theft. C2PA applies the principle to media, cryptographically signing content at creation so provenance is provable rather than inspected. Continuous device attestation extends proof beyond login, so the same device and key are present throughout a transaction. Casey noted his product held up under real pressure: “We’ve had a couple of run-ins with state actors and our product performed.”
What the shift looks like
NIST’s Cybersecurity Framework 2.0 and SP 800-63 Revision 4 both push toward phishing-resistant authentication. For security leaders evaluating their deepfake defenses, the practical steps are direct:
- Replace biometric-only verification with cryptographic passkeys. Any workflow relying solely on face or voice recognition is now a deepfake target.
- Require device-bound authentication for high-value transactions. Hardware-backed keys can’t be cloned, even if a face and voice can.
- Evaluate C2PA for media-dependent decisions — financial, legal, or operational decisions made on video or audio.
- Treat detection as a signal, not a gate. Detection can flag content for review, but it shouldn’t be the sole control.
The organizations that keep investing exclusively in detection are playing a game where the adversary improves faster than the defender. This is the same logic behind treating impersonation and social engineering as a counter-espionage problem rather than a spot-the-fake exercise: make the identity layer cryptographic rather than biometric, and deepfakes stop mattering.