FACT: A Forensic Agent with Compiled Tool-Use Trajectories for AI-Generated Image Detection

Detecting AI-generated images is only half the task: a deployed detector must also justify its verdict, yet existing detectors inherit three failure modes from their training data: real and fake images collected from different sources invite provenance shortcuts, supervised explanation corpora teach templated rationales, and a static forgery corpus leaves the decision boundary standing still while generators keep moving. We introduce \methodname{}, an adversarial reinforcement learning framework that pits two heterogeneous models against each other. A diffusion image editor learns to edit real photographs into fake counterparts of those same photographs that fool the current detector, while a reasoning MLLM learns to expose them with a verdict grounded in free-form reasoning. Both rewards are shortcut-proof by design: the attacker is credited only when its edit is faithfully executed, and the defender only when its verdict is correct. As the two models alternate, each round's attacker regenerates a harder training pool aimed at the current detector's blind spots, so the detector must generalize rather than memorize any fixed artifact distribution. Although the explanation is never rewarded, its quality rises round over round as a side effect of accuracy-only training. A detector trained within this loop improves monotonically across rounds on each of three external benchmarks.

Detectors of AI-generated images are typically trained using samples from all Generative AI architectures they must catch, and struggle as soon as a new architecture emerges. Recent approaches have explored self-supervised pre-training as an alternative solution, yet standard frameworks work against the forensic task, e.g., their augmentations overwrite the micro-statistics of image formation. This paper introduces Forensic Twins, a Self-Supervised Residual Learning (SSRL) framework whose pretext task suppresses macroscopic content availability. Each image is mapped through a frozen, off-the-shelf forensic residual extractor, from which two spatially disjoint crops are drawn. Sharing no pixel, the two views retain minimal semantic structure to align, leaving a redundancy-reduction objective with a predominant common signal: the stationary fingerprint of the image acquisition pipeline. Additionally, Forensic Twins is trained exclusively on real images; no AI-generated image is observed at any stage. Experiments show that Forensic Twins attributes AI generator sources with 56.61% accuracy, i.e., 6.13% above the previous state-of-the-art zero-shot method at 375x lower latency. We also demonstrate that fitting a Gaussian Mixture Model (GMM) offline using only the real image embeddings extracted from Forensic Twins turns it into a state-of-the-art zero-shot detector, reaching 97.99% AUC across 27 unseen AI generators, including GANs, diffusion models and commercial systems. Code, weights and exact splits will be made publicly available

As AI generated images are getting popular these days, discussions on such generations are widely existed in current public forums, software repositories, and photography tutorials. Moreover, with the help of LLM-assisted code translation, malicious actors can adopt such information to evade detectors without specialized knowledge or local GPU execution of the resulting tools. To evaluate such risks, we introduce WildFakeToolBench, a benchmark suite of $156$ executable image-processing tools derived from $220$ public sources across $142$ domains. Each tool retains its source evidence and passes syntax, execution, and output-distortion checks without using feedback from the evaluated detectors. On $105$ generated images from five generators, we evaluate six state-of-the-art detectors with images processed by the tools in our collected bench to assess their effectiveness against such in-the-wild risks. The strongest individual tools achieve attack success rates of $25.25\%$--$100\%$ across detectors, exceeding the strongest of six selected image-processing baselines for each detector. Evaluating subsets of an eight-tool pool yields worst-case coverage above $80\%$ for every detector. Effectiveness is sparse and detector-specific: high coverage across candidates does not imply reliable one-shot evasion. These findings not only highlight the potential of such web-sourced risks, but also show the potential on establishing AI-image detector robustness evaluation framework based on those in-the-wild information.