Claude Fable 5.2 Refused Every Security X-ray in Ultralytics Test

Every IEDXray request was blocked

Ultralytics tested Claude Fable 5.2 on UL37, a benchmark containing 360 images from 36 object detection datasets. The IEDXray dataset contains baggage-scanner images designed to train threat detectors for airports, borders and event security. Its 11 classes include everyday electronics, modified versions of those devices and explosives.

Claude Fable 5.2 refused all 10 IEDXray scans. In one request, Ultralytics asked the model to draw bounding boxes around a laptop and batteries visible in a bag. Anthropic's API returned: “This request triggered restrictions on content related to the development of high-yield explosives and was blocked under Anthropic's Usage Policy.”

Ultralytics said five of the 10 images contained only an ordinary phone, laptop, pager or walkie-talkie. The company said the responses had a refusal stop reason, zero output tokens and no boxes. On Ultralytics Platform, users see a message saying the process stopped before finishing its detections and suggesting another model.

A policy change from Claude Fable 5.1

Ultralytics said the block appeared to be triggered by the task and dataset context rather than by the contents of each image. Claude Fable 5.1 refused one of the same 10 scans, while Claude Fable 5.2 refused all 10.

Anthropic's launch announcement said Claude Fable 5.2's cyber and biology safeguards produced about 55% fewer unnecessary blocks than Claude Fable 5.1 in early testing. Ultralytics reported the opposite result for its security X-ray test. Anthropic also said Claude Fable 5.2 and Claude Mythos 5.2 are the same underlying model, while Mythos 5.2 ships without certain cybersecurity and biology safeguards and is available only to organizations Anthropic has verified.

According to Ultralytics, no other model tested refused an IEDXray scan. This included models from OpenAI, Google, xAI, Meta, Chinese labs and Anthropic's Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5. The 19 models hosted on Ultralytics Platform's own servers also labeled every scan.

Strong results outside security X-rays

Claude Fable 5.2 scored 0.544 mAP 50 on UL37, placing second among 44 models behind GPT-6.1 Terra. It ranked first on the stricter mAP 50-95 measure with a score of 0.347. Ultralytics said the model's boxes were the tightest it had measured.

The model finished in the top five on 22 of the 36 datasets and ranked first on UEC school lunch trays, basketball and hoop detection, and PlantDoc leaf disease. It also ranked first in retail, and narrowly led Claude Opus 5.5 in nature and agriculture. On 10 COCO 2017 reference images, it reached 0.686, the best result from a language model in that test.

Its UL37 scorecard lists precision and recall of 0.602 and 0.620, a median time of 4.6 seconds per image, and a provider cost of $37.55 per 1,000 images. There were 11 failed requests out of 360, including the 10 refusals and one request that exceeded Ultralytics Platform's 115-second limit.

What it means for annotation teams

Ultralytics said Claude Fable 5.2 is available on Ultralytics Platform today. For general-purpose labeling, its benchmark results place it among the platform's strongest tested models. For security X-ray workflows, however, the refusals mean users would need to choose another model or draw the boxes themselves.

Ultralytics reported that Moondream 3.1 scored 0.27 mAP 50 on IEDXray, while Gemini 4 Argon scored 0.30. The company said open-weights models can avoid a vendor policy layer between the user and the model's output. It also said that, after labeling, teams can train their own Ultralytics YOLO26 detector, leaving the deployed model under their control.