Catching fakes is becoming an agent's job too
A forensic agent investigates suspect video the way an analyst would, a world model spots fakes it was never trained on, and two vendors claim near-perfect scores on the same audio benchmark. Detection is no longer a single classifier.
- In a preprint, the WeVisForensics agent achieves 95.2% detection accuracy across in-domain, out-of-domain, and agentic benchmarks. Self-reported m:…7415
- In a preprint, the training-free SaySee detector achieves a mean AUC of 98.2% across 24 evaluation sets, beating the best existing detector by 4.2 points. Self-reported m:…3957
- The company says its Arcas technology ranked #1 in the Podonos audio deepfake benchmark with 99.93% accuracy, making 3 errors across 4,524 clips. Self-reported m:…0528
- The company says its DETECT-World model achieved 99.5% pooled accuracy on the Podonos benchmark, with a 0.4% false-negative rate. Self-reported m:…0070
- New-account fraud affected 5.4 million US victims in 2025, a 31% increase on the previous year. m:…0070
An agent that investigates
In a preprint, WeVisForensics achieves 95.2% detection accuracy across in-domain, out-of-domain, and agentic benchmarks. The system is a hierarchical agentic framework rather than a single classifier: dual-timescale policy optimization improves forensic capabilities by closing the loop between investigator and inspector agents, and it releases ForenCue-30K, a dataset containing forensic annotations and interaction trajectories for training forensic agents. The idea is that detecting a fake video is an investigation, with evidence gathered, checked, and re-checked, not a single yes-or-no score. m:…7415
That shift matters because the fakes keep changing. A detector trained to spot one generator's fingerprints fails on the next generator, so the systems that survive are the ones that reason about what is in the video rather than memorising what last month's fakes looked like. An agent that can look again, from a different angle, has a better chance than a classifier that gets one pass. Both of the preprints in the next section agree on the diagnosis even though their methods differ: a detector that memorises generator fingerprints is betting that the next fake will look like the last one, and the evidence says that bet loses. m:…7415 m:…3957
Catching fakes without ever seeing one
A second preprint takes the idea further by asking whether a world model can predict a deepfake video. Manipulated video clips produce, on average, larger prediction residuals than authentic clips when compared against a natural video predictive model, and a model trained only on FF++ achieves 90.5 AUC on Celeb-DFv2 using the consistency-margin prior. The direction of the consistency margin affects cross-dataset generalization, which is the point: the detector learns what real video should look like, and flags whatever does not behave like reality. m:…3947
A third preprint makes the same bet for images, without training on fakes at all. Training on generated images limits detector generalization and robustness due to generator-specific traces and format and content bias, so SaySee is a training-free detector that avoids the problem entirely. It achieves a mean AUC of 98.2% across 24 evaluation sets and outperforms the best existing detector by 4.2 points in mean AUC. The lesson of both papers is the same: the less a detector depends on examples of yesterday's fakes, the better it handles tomorrow's. m:…3957
The commercial race, measured on the same yardstick
The vendors are racing on the same ground. Neural Defend ranked #1 in Podonos' independent Audio Deepfake Detection Benchmark, the company says, which evaluated 28 systems, and achieved 99.93% detection accuracy, with two Arcas variants taking both the #1 and #2 positions and each making 3 errors across 4,524 English studio-audio clips. The company is building multimodal AI detection technology across audio, video, images, and digital identity verification, so the audio result is presented as one part of a wider detection stack. m:…0528
Resemble AI reports on the same benchmark and the same 4,524-clip test set: its DETECT-World model achieved 99.5% pooled accuracy on the Podonos Deepfake Detection Benchmark, with a 0.4% false-negative rate and a 0.12 real-time factor across 4,524 audio files, and its detection models support more than 50 languages and can be deployed within customer-controlled infrastructure. The commercial context is urgent, with new-account fraud affecting 5.4 million US victims in 2025, a 31% increase on the previous year. Real-time detection systems face challenges including latency constraints, false positives, and integration complexity, which is why the real-time factor and the false-negative rate matter as much as the headline accuracy. m:…0070
What is still unresolved
- Do benchmark scores on studio audio hold up on noisy, real-world phone calls and compressed video?
- Will training-free and world-model detectors keep their edge as generators learn to mimic natural video dynamics?
- How do buyers compare vendors who report different metrics on the same benchmark?
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.