A NIRCam-dark Galaxy Detected with the MIRI/F1000W Filter in the MIDIS/JADES Hubble Ultra Deep Field

We report the discovery of Cerberus, an extremely red object detected with the MIRI Deep Imaging Survey (MIDIS) observations in the F1000W filter of the Hubble Ultra Deep Field. The object is detected at signal-to-noise ratio (S/N) ∼ 6, with F1000W ∼ 27 mag, and undetected in the NIRCam data gathered by the JWST Advanced Deep Extragalactic Survey (JADES), fainter than the 30.0–30.5 mag 5σ detection limits in individual bands, as well as in the MIDIS F560W ultradeep data (∼29 mag, 5σ). Analyzing the spectral energy distribution built with low-S/N (<5) measurements in individual optical-to-mid-infrared filters and higher-S/N (≳5) measurements in stacked NIRCam data, we discuss the possible nature of this red NIRCam-dark source using a battery of codes. We discard the possibility of Cerberus being a solar system body based on the <0.″016 proper motion in the 1 yr apart JADES and MIDIS observations. A substellar Galactic nature is deemed unlikely, given that the Cerberus’s relatively flat NIRCam-to-NIRCam and very red NIRCam-to-MIRI flux ratios are not consistent with any brown dwarf model. The extragalactic nature of Cerberus offers three possibilities: (1) a z ∼ 0.4 galaxy with strong emission from polycyclic aromatic hydrocarbons—the very low inferred stellar mass, M ⋆ = 105–106 M ⊙, makes this possibility highly improbable; (2) a dusty galaxy at z ∼ 4 with an inferred stellar mass M ⋆ ∼ 108 M ⊙; and (3) a galaxy with observational properties similar to those of the reddest little red dots discovered around z ∼ 7, but Cerberus lying at z ∼ 15, with the rest-frame optical dominated by emission from a dusty torus or a dusty starburst.


INTRODUCTION
Building increasingly powerful telescopes operating at redder and redder wavelengths and (unexpectedly) discovering new types of galaxies at higher and higher redshifts is a classical industry now, established nearly 40 years ago.Indeed, deep optical surveys carried out in the late twentieth century using highly sensitive detectors were found to miss interesting galaxy populations of evolved and dust-rich galaxies at cosmological distances, both types being relatively bright at near-infrared wavelengths and easily detectable by new instruments, even though they were significantly less efficient in detecting photons and were affected by much larger backgrounds.
The key for their discovery was large (red) colors (flux density ratios typically larger than an order of magnitude) between a near-infrared and an optical band, i.e., they were extremely red objects.
The first Extremely Red Objects (EROs) were reported using new ground-based near-infrared telescopes built nearly 40 years ago.They were found to be members of cosmologically relevant galaxy populations, comprising a variety of galaxies ranging from high redshift quasars to very young and dust-rich galaxies, or even "high" redshift passive galaxies (see, among many, Elston et al. 1988;McCarthy 2004;Graham & Dey 1996;Franx et al. 2003;Pozzetti & Mannucci 2000).
With the development of very sensitive instruments onboard space telescopes such as Spitzer, the search was moved to longer wavelengths in the mid-infrared, at first compared with optical datasets (Wilson et al. 2004;Yan et al. 2004).
Remarkably, when the Hubble Space Telescope gathered extremely deep data taken in the near-infrared up to 1.6 µm, with surveys such as CANDELS or the WFC3 Hubble Ultra Deep Field (Koekemoer et al. 2011;Grogin et al. 2011;Koekemoer et al. 2013), Spitzer/IRAC data were still shown to detect unique distinctively red objects which were only visible at wavelengths longer than 3 µm, or at least were extremely faint in the near-infrared and optical spectral ranges.These socalled HST-dark galaxies were studied for around a decade, with significant uncertainties about their nature due to the strong sensitivity limitations of our multi-wavelength telescopes, even those operating in very different observational windows such as the (sub-)millimeter range (see, e.g., Simpson et al. 2014;Franco et al. 2018;Alcalde Pampliega et al. 2019;Wang et al. 2019).JWST has started, from the earliest phases of the mission, to shed light on those HST-dark galaxies, i.e., mid-infrared bright near-infrared/optically faint sources which now constitute the bulk of the JWST galaxy exploration (Pérez-González et al. 2023a;Nelson et al. 2023;Barrufet et al. 2023;Williams et al. 2023a;Gillman et al. 2023).
Last but not least, EROs, especially if selected at the faintest magnitudes and with appropriate bands for the color, overlap with very high-redshift galaxy samples selected using the Lyman break technique (Steidel et al. 1996) at higher and higher redshifts (e.g., Bouwens et al. 2011).
Remarkably, with the advent of JWST, both industries, searching for red objects and for those with higher and higher redshifts, have been merged at a profound level, with the possibility of finding double-break (Lyman and Balmer) red galaxies (Labbé et al. 2023a), now known as little red dots (Furtak et al. 2023;Übler et al. 2023;Matthee et al. 2023;Barro et al. 2023;Greene et al. 2023;Williams et al. 2023a;Killi et al. 2023;Pérez-González et al. 2024;Kokorev et al. 2024), or high redshift galaxies with very high equivalent width emission lines, first discovered with Spitzer (Smit et al. 2014), now appearing in many JWST works (Pérez-González et al. 2023a;Rinaldi et al. 2023).Interestingly, some of these sources have been confirmed to present broad emission line components possibly linked to an active galactic nucleus (AGN; Kocevski et al. 2023;Greene et al. 2023).
The interesting subtopics of finding and characterizing dusty galaxies on the one hand, and high redshift galaxy candidates on the other, has been demonstrated to be very relevant for JWST surveys, which can easily misidentify one type of galaxy with the other (see, e.g., Zavala et al. 2023;Naidu et al. 2022).
The MIRI instrument onboard the JWST (Rieke et al. 2015;Wright et al. 2023) now offers the possibility to search for red galaxies with longer color baselines, beyond the wavelengths first probed by Spitzer/IRAC and now easily accessible with impressive depths (∼ 30 mag and beyond) using JWST/NIRCam.Even though MIRI cannot reach NIRCam depths, very deep imaging at 5.6 -12.8 µm reaching magnitudes around 27 -29 is already available and opens a whole new window for the search and characterization of high redshift galaxies (Papovich et al. 2023;Barro et al. 2023;Rinaldi et al. 2023;Pérez-González et al. 2024).
In this paper, we report the discovery of a MIRI-bright red object appearing in the MIRI Deep Imaging Survey (MIDIS) F1000W imaging of the Hubble Ultra Deep Field ( Östlin et al., 2024, in prep.).The source is extremely faint at NIRCam long wavelengths, i.e., it is a NIRCam-dark object lying at the edge of the detection limit (∼ 30.5 mag) of one of the deepest JWST surveys available to date, JADES (Eisenstein et al. 2023a), having not been catalogued before (Rieke et al. 2023;Eisenstein et al. 2023b).Consequently, our object presents an extremely red color between 4.4 and 10 µm, and, in fact, the source is also red with respect to the bluest MIRI filter, as revealed by the MIDIS F560W imaging.
This letter is organized as follows.Section 2 presents the deep MIRI data used in this letter, as well as ancillary observations taken with NIRCam, ALMA, and VLT.Section 3 describes our method to detect NIRCamdark galaxies, and the case for a positive detection in the Hubble Ultra Deep Field, the source we call Cerberus.Section 4 presents the spectral energy distribution of Cerberus and discusses the several possible interpretations, analyzing the implied physical properties for each one of them.Finally, Section 5 presents a summary of our findings.
Throughout the paper, we assume a flat cosmology with Ω M = 0.3, Ω Λ = 0.7, and a Hubble constant H 0 = 70 km s −1 Mpc −1 .We use AB magnitudes (Oke & Gunn 1983).All stellar mass and SFR estimations assume a universal Chabrier (2003) initial mass function (IMF), unless stated otherwise.(MIDIS, Östlin et al. 2024, in prep.;PID 1283) in December 2023.The observation consisted of 11 exposures, each with 100 groups, FASTR1 readout, and 10 integrations, for a total on source exposure time of 30800 s, centered on the Hubble Ultra Deep Field (HUDF).The dithering pattern was set to large-size cycling, with the 11 exposures taken in different positions on the sky separated by up to 10 ′′ .
The reduction of the MIRI data followed the methodology described in Pérez-González et al. (2024), based on the official jwst pipeline (v1.13.4,jwst 1202.pmap) and offline bespoke procedures.Briefly, a super-background strategy is used to build background maps for each single image using all the other exposures (since they were taken during the same campaign), which results in a very homogeneous background in terms of level and noise.Quantitatively for the data used in this letter, the standard deviation of the background pixels for each cal.fits file obtained with the standard official pipeline is reduced by a factor of ∼ 3.5 by using our superbackground method, and the depth in the final mosaic improves by 0.8 mag.We refer the reader to Appendix A in Pérez-González et al. (2024) for more details.Known sources are masked to avoid biasing the determination of the very local background in the super-background image.Our F1000W final mosaic, reduced with a 60 milliarcsec pixels, presents an average 5σ depth of 26.8 mag, measured in a 0. ′′ 4 radius circular aperture, and taking into account noise correlation (linked to the drizzling method used in the mosaicking) and after applying the aperture correction (a 1.3 factor).
We present the MIDIS F1000W data in Figure 1 with an RGB image built in combination with some NIRCam long wavelength data.

Ancillary data
In this subsection, we describe other datasets in the HUDF that we have used to complement the MIRI data.

Other MIRI data
For this letter, we combined the recently acquired MIDIS F1000W data with the F560W ultra-deep observations also carried out by MIDIS in December 2022.These data will be described in Östlin et al. (2024, in prep.) and consist of ∼ 40 hours on source taken in the HUDF, which allowed us to reach 28.7 mag 5σ for point-like sources measured in an r=0.′′ 23 circular aper- In the center of the figure, we show an RGB composition of the MIDIS F1000W field.The color images have been built with JADES data in 2 NIRCam filters, F277W and F356W, and the MIDIS MIRI F1000W filter (all images convolved to the same PSF of MIRI/F1000W).In the background, we show the HUDF JADES data in grayscale.We provide a series of zoomed-in RGB frames that lead to the NIRCam-dark MIRI source, Cerberus, whose image is located at the bottom right corner.The NIRCam-dark MIRI source Cerberus is highlighted with a square in that RGB frame.Additionally, we provide postage stamp images (2 ′′ ×2 ′′ ) of Cerberus in MIRI/F1000W, MIRI/F560W, and NIRCam/F444W (left column), showcasing that Cerberus (marked with an r=0.′′ 25 green circle) is clearly detected only at 10 µm.
ture (the radius being chosen to ensure a ∼ 70% encircled energy).

NIRCam data
The MIRI/F1000W data were also complemented with NIRCam imaging taken by the JWST Advanced Deep Extragalactic Survey, JADES (Eisenstein et al. 2023a), Data Release 2 (Eisenstein et al. 2023b), which includes observations from the JWST Extragalactic Medium-band Survey (JEMS, Williams et al. 2023b) and the First Reionization Epoch Spectroscopically Complete Observations (FRESCO, Oesch et al. 2023).This dataset provides a total of 14 bands from 0.9 to 4.8 µm (6 at short-wavelength, SW, and 8 at longwavelength, LW), with 5σ depths ranging from 30.5 to 30.9 mag (measured in a 0. ′′ 2 radius circular apertures).We remark that JADES is one of the deepest NIRCam surveys on the sky, only matched in depth (in some bands) by the MIDIS NIRCam-parallel project (Pérez-González et al. 2023b) and The Next Generation Deep Extragalactic Exploratory Public Near-Infrared Slitless Survey, NGDEEP (Bagley et al. 2023), which means that our search for NIRCam-dark galaxies is really probing the current depth limitations of JWST extragalactic surveys.

MUSE data
The HUDF has been extensively studied in the last decade with the Multi Unit Spectroscopic Explorer (MUSE, Bacon et al. 2010) mounted on the Very Large Telescope (VLT) as part of the MOSAIC and UDF-10 fields (GTO programs 094.A-0289(B), 095.A-0010(A), 096.A-0045(A) and 096.A-0045(B), PI: R. Bacon) and the most recent MXDF observations (GTO Large Program 1101.A-0127, PI: R. Bacon).The MUSE data has a spectral wavelength range between 4700 -9300 Å and a spectral resolving power (R) that ranges from 1770 (4800 Å) to 3590 (9300 Å).The HUDF dataset was taken during 140 hours on-source with the VLT Adaptive Optics Facility (AOF) and its GALACSI adaptive optics module, as detailed in Kolb et al. (2016) and Madec et al. (2018), achieving a spatial sampling of 0. ′′ 2 × 0. ′′ 2, with strong variations of up to a factor of ×4.More details about these programs can be found in Bacon et al. (2017Bacon et al. ( , 2023)).

SELECTION AND VALIDATION OF NIRCAM-DARK SOURCES
This letter presents a source that was discovered on the MIDIS F1000W image and found not to be included in existing NIRCam published catalogs (Rieke et al. 2023;Eisenstein et al. 2023b).In this section, we describe the selection procedure for such types of NIRCamdark sources.

Selection of NIRCam-dark sources
In order to select NIRCam-dark sources, we executed the sextractor software (Bertin & Arnouts 1996) directly using the MIRI/F1000W as the detection image.We followed the hot-mode extraction described in Galametz et al. (2013), which is particularly effective for detecting extremely faint sources.
Once we had a hot MIRI/F1000W catalog, we anti cross-matched it (using a 0. ′′ 3 search radius) with the official JADES DR2 catalog published by Eisenstein et al. (2023b, which superseded the DR1 catalog), constructed by selecting sources in a LW stack.The goal of the anti cross-match was to isolate sources uniquely detected by MIRI/F1000W.The result was a list of 17 potential NIRCam-dark sources selected at 10 µm.
We meticulously examined each of the potential NIRCam-dark sources in the final mosaic to rule out the presence of an unflagged cosmic ray or any kind of hot and/or bad pixels that could have caused a spurious detection.
Visual inspection of the individual exposures revealed that for 16 of the 17 candidates, there was indeed a hot pixel in one of the 11 exposures which had survived the flagging and stacking process performed to produce the final mosaic.We automated and generalized the source validation methodology by producing stacked mosaics with a limited number of exposures, excluding 1, 2, or 3 of the full set randomly.This procedure, obviously, affects the depth.Explaining the latter case in more detail as an example, we started from our collection of 11 * cal.fits files, and we created the 165 unique mosaics which can be obtained by excluding sets of 3 * cal.fits files for each partial mosaic.After building the 165 realizations, we inspected the 17 potential NIRCam-dark sources, concluding that only one source consistently appeared in all 165 different partial mosaics.The remaining candidates were found to vanish in some of the realizations, in fact even in the mosaics built excluding one single exposure, indicating their most-likely spurious nature.To further validate these findings, we made an in-depth analysis of each * cal.fits file, confirming the presence of unflagged hot and/or bad pixels in some of them that indeed led to false detections in a number of the realizations.
One source of the 17 pre-selected candidates did not vanish in any of the partial mosaics, even when removing 3 random images.For this source, we measured photometry in a small 0. ′′ 25 radius aperture (which maximized S/N on the partial mosaics), as well as the sky noise in a square 3 ′′ ×3 ′′ region around, on all the mosaics produced by excluding 1, 2, and 3 exposures.We found consistent photometry and a degradation of the S/N consistent with the differences in exposure time.For example, for the skip-3 mosaics, we found an average magnitude 27.05±0.19mag, consistent with our measurement in the full mosaic (see Section 4).
However, we found 2 contiguous exposures, numbers 7 and 8, which, if removed, produced a dimming of the source to 27.4 mag, still within 2σ of the measurement using the full dataset.Each one of these exposures has an integration time of 2800 s, with the start of the first exposure and the end of the second separated by ∼5700 s.We examined these exposures in detail and found no signs of a cosmic ray hit, but a slightly enhanced signal in an extended region.This could be due to the general properties of the detector (which consistently showed a dimmer region in the super-background frames for all images).It could also be due to a cosmic ray shower, which are known to affect MIRI imaging observations (Dicken et al. 2024), with an incidence of several per exposure.Cosmic ray showers are extended persistence artifacts from cosmic ray hits and typically affect the detector for several minutes (Wright et al. 2023).However, a high flux cosmic ray hit could cause a shower lasting long enough to enhance the signal seen in these exposures.We also note that the enhancement in signal is stronger in exposure 7 compared to 8, which is also consistent with persistence (see Álvarez-Márquez et al. 2023).Given that the two relatively (and slightly) brighter images were contiguous, we do not rule out a transient behavior of the source.Based on the perseverance of the signal from the source in all partial stacks Note that the spurious pixels in the F460M image (see text for details) are removed in the median stacking procedure.The angular size of the images and zoomed inset is 3 ′′ ×3 ′′ and 0. ′′ 6 ×0.′′ 6, respectively.The red (blue) circle shows the location of the detected source in F1000W and has a radius of 0. ′′ 2. The green filled circle shows the resolution of the F1000W image (with a diameter of 0. ′′ 328).In the zoomed insets, we also show the centroid positions with the blue and red crosses, corresponding to the F1000W image and the NIRCam stack, respectively.
of the MIRI F1000W dataset, we concluded that it was not spurious.
The NIRCam-dark source we identified with this methodology was dubbed Cerberus, to acknowledge that it is a weird beast lurking in the darkest corners of the JWST surveys and with different interpretations/heads for its nature, as will be shown in Section 4. Its coordinates (J2000) are α = 03:32:38.09(hours), δ =−27:47:11.81(degrees).NIRCam and MIRI images of Cerberus are presented in Figure 1.

Multi-wavelength validation and observational properties of the NIRCam-dark object Cerberus
In order to further validate and characterize Cerberus, we searched for detections by forcing photometric measurements in the datasets described in Section 2, more specifically in the ultra-deep observations (among or even the deepest on the sky) provided by JWST/NIRCam, ALMA, and VLT/MUSE.

JWST/NIRCam measurements
Directly linked to our selection criteria, the NIRCamdark Cerberus object was not present in any of the catalogs of the JADES NIRCam survey (Rieke et al. 2023;Eisenstein et al. 2023b).We then proceeded to force photometric measurements in all bands (described in Section 4).Forced photometric measurements in all NIRCam individual bands do not provide any measurement signal-to-noise ratio S/N > 5.
We then stacked NIRCam images to search for faint emission.We constructed two different sets of stacks.One set consisted in cumulative stacks constructed by averaging NIRCam bands starting from the reddest and from the bluest.Another set consisted of disjoint stacks (to avoid noise spikes in one band affecting the measurements in cumulative stacks), combining 2, 3, 4 and 5 bands, always starting from the reddest and not repeating them in different stacks of the same set.
Figure 2 shows a stacked NIRCam image (including all filters at λ > 3µm) convolved to the resolution of the F1000W data (FWHM=0.′′ 328).Each calibrated individual NIRCam filter was rebinned to a 0. ′′ 06 pixel scale and PSF-matched to F1000W using WebbPSF (v.1.2.1 Perrin et al. 2014) models and the method described in Melinder et al. (2023), and then median-stacked.We measured the centroid of the source for both JWST instruments (using photutils with a circular footprint of 3 pixels radius), finding an agreement within 0.26 pixels, i.e., 0. ′′ 016.Remarkably, the source is slightly extended to the NW in both images.The morphology of the source is similar for the NIRCam stack and the MIRI F1000W image, with a tadpole-like shape extending to the upper right, confirming a detection of the source with NIRCam stacked data (see next section).We also notice a qualitative resemblance between the structure of Cerberus and the z = 10.6 galaxy GN-z11 (Oesch et al. 2016;Tacchella et al. 2023), with a concentrated flux distribution identified with the core of the galaxy and an extended haze component.

MUSE measurements
We extracted the MUSE 1D spectrum at the coordinates of our target within circular apertures of different sizes (0. ′′ 4 and 0. ′′ 8 radius) but found no significant spectral feature at the position of the NIRCam-dark candidate.This is also confirmed by the publicly available catalog of detected sources by Bacon et al. (2023).

ALMA measurements
The source is not detected at 1.2 mm and 3.0 mm.This finding implies 5σ upper limits on the flux density at 1.2 mm and 3.0 mm of 47 µJy and 7 µJy respectively, assuming it is effectively an unresolved point-like source.

Multi-wavelength Photometry
A spectral energy distribution (SED) for Cerberus was built by measuring photometry in all available NIRCam (including both individual bands and stacks) and MIRI data, forcing measurements at the position determined for Cerberus with the MIRI data.Given the faint and small nature of our source, we measured forced photometry in several circular apertures with radii ranging from 0.1 ′′ to 0.6 ′′ .We measured the background in a small square region around Cerberus (3 ′′ on the side) and calculated the noise using non-adjacent pixels (5 pixels apart) to take into account noise correlation, as explained in Pérez-González et al. (2023b), obtaining similar depth results (within 0.2 mag) to those reported in Eisenstein et al. (2023a) for the NIRCam bands.Photometric measurements are provided in Table 1.
None of the measurements in individual NIRCam bands have a > 5σ significance, which we use as our limit to define a detection in each filter.
These measurements indicate that Cerberus is not robustly detected in any individual NIRCam filter, but consistent with being slightly below the 5σ detections limits in some bands, quoted to be around 30-31 mag for the JADES DR2 data (Eisenstein et al. 2023a).We remark that the F460M filter, which provides the brightest magnitude, presents a group of 4 pixels to the North of the position of our galaxy, which we confirmed to be an artifact of the reduction, and it is not seen in any other band (even the ones with deeper limiting magnitudes).We excluded this filter from the rest of analysis.
We measured photometry in the stacked data in the same apertures used for the individual bands, assigning average aperture corrections for the filters included in each stack.The 5 LW stacks with the highest number of bands all provide S/N > 5 fluxes (the average being S/N ∼ 6), ranging from 29.9 to 30.8 mag.This means that the (low-S/N) SED is characterized by a red color extending through all the S/N > 5 stacks from 2 to 4 µm and beyond, up to 29.9±0.2mag for the first (reddest) stack (including F480M+F444W), which reaches ∼ 5 µm.For the SW stacks, no significant signal is recovered in any of them.We measure a 5σ upper limit of ∼ 31 mag for the stack adding all SW data.The comparison between SW and LW stacks implies a ≳0.6 mag color jump from the ∼ 1 − 2 to the ∼ 3 − 5 µm ranges.
The SED of Cerberus is, in fact, very similar to the little red dots (LRDs) recently presented and characterized in a number of papers (e.g., Labbé et al. 2023a,b, Matthee et al. 2023, Barro et al. 2023, Williams et al. 2023a, Pérez-González et al. 2024, Kokorev et al. 2024), but shifted to longer wavelengths.Based on detections of Cerberus in stacked images, we can infer that the F277W − F444W > 1 mag general selection criterion of LRDs, and, more specifically, the 3 mag color of the example source 203749 in Pérez-González et al. (2024), compares well, if we multiply the observed wavelengths by a factor of ∼2, with the F444W − F1000W ∼ 3.5 mag, or, using the F560W 5σ upper limit, F560W − F1000W ≳ 2 mag.The flat F150W−F200W color distinctive of LRDs (the largest value allowed by the selection is typically 0.5 mag), i.e., a change in slope at around 0.3-0.4µm in the rest-frame for the typical redshift of LRDs 5 < z < 9, compares well with the color between 3 and 4 µm for Cerberus, just multiplying again the observed wavelengths by a factor of ∼2.We remark that this comparison is based on S/N>5 measurements in stacked data.That rest-frame UV color is indeed consistent with the reddest LRDs in Pérez-González et al. (2024) or the LRD of which a NIR-Spec spectrum was presented in Killi et al. (2023).The analogy can also be established in terms of the apparent magnitude: the typical magnitude of LRDs in the blue part of their SED lies around 27-28 mag, Cerberus being ≳ 10 times dimmer in NIRCam LW bands.We note that the difference in luminosity distance between z ∼ 7, the median redshift of LRDs (Pérez-González et al. 2024) and z ∼ 14 (a factor of 2 larger, see also Section 4.2.1) will translate to a flux ratio between bands covering the same rest-frame wavelength around an order of magnitude for the same type of object.

Analysis of the SED of Cerberus
In the following subsections, we discuss possible interpretations of the SED of Cerberus.The NIRCam and MIRI datasets were taken more than one year apart.However, given the spatial coincidence of the object in the NIRCam and MIRI observations (centroids separated by 0. ′′ 016), we proceed by assuming that they are the same object and not a chance alignment of two different objects.

Extragalactic origin
The SED of Cerberus was fitted to estimate a photometric redshift and obtain its physical properties in a 2-step procedure.We used several codes for this purpose, separating the two tasks to analyze facets of each method and understand the (possibly systematic) uncertainties in detail.We used the fluxes described above as well as the ALMA upper limits quoted in Section 2, shown to be important for high redshift sources (even for determining their redshift, see Meyer et al. 2024), and particularly for an ERO such as Cerberus.We provide details about the fitting methods in Appendix A.
The results of our SED modeling analysis are given in Figures 3 (photometric redshift probability distributions) and 4 (SED models).
We remark that the SED of Cerberus is very noisy, given that it is an ultra-faint source at the very depth edge of the JWST surveys.Therefore, the physical properties of the source are highly uncertain, starting from its redshift (see several peaks in Figure 3).
However, leaving the redshift apart, there are two physical properties for which our results are similar among all SED modelling codes, mainly because they are linked to the 2 main observational characteristics of Cerberus, i.e., its faintness (translating to relatively low luminosity) and red color.The very red color of our NIRCam-dark object can only be reproduced with significant amounts of dust, either linked to PAHs if the redshift is low, z ∼ 0.4, or warm/hot dust if the redshift is around z ∼ 4 or z ∼ 15.For the lowest redshift, all codes provide very low stellar masses, around ∼ 10 5−6 M ⊙ .The very small stellar mass, relatively high dust attenuation, and strong PAH emission is an uncommon combination, at odds with the massmetallicity relationship (e.g., Tremonti et al. 2004) and the weakening of PAH features at low metallicities (e.g., Engelbracht et al. 2005).For the higher redshift solutions, z ∼ 4 and z ∼ 15, all codes provide masses around ∼ 10 7−8 M ⊙ , and they all need large warm dust contents, which can be explained with a very compact starburst or an AGN torus (similar to those detected at high redshift by JWST, see Maiolino et al. 2023, as well as Pacucci et al. 2023;Pacucci &Loeb 2024, andalso Hamann et al. 2017).

Solar System origin
A bright mid-infrared object such as Cerberus matches the expected properties of Solar System bodies, e.g., asteroids.Given the long observational MIRI/F1000W campaign, we would expect to see proper motion, which is not the case, unless the asteroid is distant and/or near aphelion.However, the fact that the NIRCam data detect Cerberus (in stacked images) and the observations from MIDIS and JADES were taken more that one year apart indicates that Cerberus cannot be a Solar System body (assuming that the NIRCam and MIRI sources are the same object and not a fortuitous superposition).In Section 3.2, we also measure the centroid shift of the object between the MIRI and NIRCam data to 0. ′′ 016 which is consistent with a zero offset.This makes the source even more unlikely to be a Solar System body.

Galactic origin
As shown by Langeroodi & Hjorth (2023a), Burgasser et al. (2024), andHainline et al. (2023) for LRDs, red sources detected by JWST in cosmological fields such as GOODS-S or A2744 can be brown dwarfs.Following the method in Langeroodi & Hjorth (2023a), we fit the photometry of Cerberus with brown dwarf atmosphere templates.We include the Sonora cloud-free models (Karalidi et al. 2021;Marley et al. 2021) as well as the ATMO2020++ cloud-free adiabat-adjusted T and Y dwarf models (Leggett et al. 2021;Meisner et al. 2023, see also Suárez & Metchev 2022).In particular the ATMO2020++ models were included to ensure coverage of extremely cool brown dwarfs with 250 K < T eff < 500K.Compared to warmer brown dwarfs, these models more closely resemble the SED shape of Cerberus, which is relatively faint at near-infrared wavelengths while getting significantly brighter at 10 µm.
The best-fit brown dwarf template is shown in the top-left panel of Figure 4. We infer a best-fit effective temperature (T eff [K]) of 300 +100 −25 , surface gravity (log g [cm s −2 ]) of 3.0 +1.0 −0.5 , and distance of 332±67 pc.Its effective temperature would most likely classify Cerberus as a Y-type dwarf (see Figure 12 in Langeroodi & Hjorth 2023a).Compared to the UltracoolSheet compilation of known brown dwarfs with parallax distances (Best et al. 2020), its inferred distance (∼ 300pc) would make Cerberus one of the farthest (top 20) brown dwarfs discovered to date (see Figure 11 in Langeroodi & Hjorth 2023a).This distance at the high Galactic latitude of Cerberus would place it at the edge of the thin disk (see Figure 6 in Hainline et al. 2023).Recently, three T-type dwarfs were discovered and spectroscopically confirmed with NIRSpec prism (Langeroodi & Hjorth 2023a;Burgasser et al. 2023) at estimated distances between 0.7 and 4.8 kpc, all most likely outside the thin disk; among them A2744-BD3 is a late T-type at 755 pc.
Our best-fit to a brown dwarf spectral template, shown in the top-left panel of Figure 4, nicely fit the F444W−F1000W color and is consistent with the nondetection in the deep MIDIS F560W data.However, that brown dwarf model is very red in both the NIRCam SW-to-LW and 3-to-4 µm colors, which is not favored by our measurements (in stacked data), pointing to a flatter SED.
In addition, both the MIRI and stacked NIRCam images (see Section 3.2) shows the source to be slightly extended and asymmetrical as compared to the PSF.The two instruments, NIRCam and MIRI, indicate a morphology for the source that makes it unlikely that it is a point source, thus disfavoring the single brown dwarf scenario.
The slight extension of order 0. ′′ 1 alone would however not argue against a brown dwarf binary.Even at only 100 pc distance, a semi-major axis of 0. ′′ 1 for a binary of two identical brown dwarfs/planets with masses of (0.0003 − 0.07) M ⊙ each would result in orbit periods of order 100-1000 years, i.e., the elongated image would not vary between the NIRCam and the MIRI observations.

SUMMARY AND CONCLUSIONS
We report the discovery of a NIRCam-dark source identified in the MIRI Deep Imaging Survey (MIDIS) of the Hubble Ultra Deep Field carried out with the F1000W filter.We call this source Cerberus.The source has a magnitude around 27 mag (6σ) in F1000W and is extremely faint in all the NIRCam data taken by the deepest survey on the sky, the JWST Advanced Deep Extragalactic Survey, JADES, implying a magnitude fainter than ∼30.5 at wavelengths ≲5 µm.Cerberus is also undetected in the MIDIS F560W ultra-deep observations of the field, i.e., F560W≳29 mag.Our analysis of MIRI F1000W mosaics produced with limited datasets as well as the detection in NIRCam stacks at ∼ 5σ level at wavelengths longer than ∼ 3 µm, both confirm Cerberus is real and qualifies as an extremely red object, with a color around F444W-F1000W ∼ 3.5 mag and F560W-F1000W ≳ 2 mag.The morphology of the source in the MIRI F1000W image matches that observed in the NIRCam stacked data (with 1 year passing between both observations), in both cases presenting a main point-like component and similar-sized (i.e., unresolved) haze, further supporting that Cerberus is a real source not linked to a fortuitous alignment.
We discuss the possible nature of Cerberus and identify 2 possible Galactic source types and 3 different kinds of galaxy that could match its spectral energy distribution.
Given that the MIRI and NIRCam observations were taken with a one year epoch difference, we discard a Solar System origin.The NIRCam-to-NIRCam and NIRCam-to-MIRI colors do not match any known substellar object, even the coolest and reddest brown dwarfs present mid-infrared colors that are not compatible with the spectral energy distribution of Cerberus.
Concerning an extragalactic origin, we identify 3 possibilities: • A low-redshift solution, z ∼ 0.4, would imply that Cerberus is a dusty galaxy with strong emission from polycyclic aromatic hydrocarbons or warm dust heated by an active galactic nucleus.The low mass obtained for this solution, M ⋆ ∼ 10 5 M ⊙ , would point to an unknown type of dwarf galaxy with large dust content and no PAH depletion.
• The second possibility is a dusty starburst or post-starburst galaxy, or a galaxy hosting a midinfrared bright obscured active galactic nucleus, at z ∼ 4 with a stellar mass M ∼ 10 8 M ⊙ and high extinction.
• The third possibility is a z ∼ 15 galaxy, the red color emanating from a spectral energy distribution dominated by emission from an obscured active nucleus with a bright torus significantly contributing to the flux in the rest-frame optical range, with shorter wavelengths dominated by a M ⋆ ∼ 10 7 M ⊙ galaxy.
Despite vetting all the imaging data meticulously, confirming beyond doubt that the source is real, we have to accept that the available data are insufficient to establish the true nature of Cerberus with certainty.Given that the source is located in the deepest HST and JWST field, deeper imaging will likely be available in the near future and help to obtain a better understanding of this unique object.In any case, further investigation of Cerberus will challenge the capabilities of JWST, as it was the case in the past for objects discovered at the limit of photometric surveys (remarkably, with spacebased missions such as HST or Spitzer), whose spectroscopic follow-up (typically with ground-based facilities) was tremendously exigent (but still attempted with varying success).The discovery and analysis presented in this letter demonstrates the feasibility of identifying NIRCam-dark sources when selected as extremely red objects with MIRI observations up to at least 10 µm, maybe revealing new types of Milky Way sub-stellar bodies or of galaxy populations more easily detectable through ultra-deep photometric surveys in (some of) the reddest wavelengths probed by JWST.Ö and AB acknowledge support from the Swedish National Space Administration (SNSA).RAM acknowledges support from the Swiss National Science Foundation (SNSF) through project grant 200020 207349.AGN emission is modeled with a QSO average spectrum (Vanden Berk et al. 2001;Glikman et al. 2006).The dust emission from the AGN is modeled with the self-consistent templates of AGN tori presented in Siebenmorgen et al. (2015), and dust emission linked to star formation is added using the models from Draine & Li (2007).CIGALE (Boquien et al. 2019) SED-fitting analysis assume an SFH modelled using a constant star-formation rate with ages ranging from 1 to 100 Myr.We adopted the stellar populations models from Bruzual & Charlot (2003) with solar metallicity, and the Chabrier (2003) IMF.We included nebular continuum and emission lines using solar metallicity, electron density of 100 cm −3 , and ionised parameter equal to log(U )=−2.The dust attenuation and far-IR emission uses the Calzetti Law (Calzetti et al. 2000) and the Draine et al. (2014) models, respectively.AGN emission is added using the Fritz et al. (2006) models following the initial parameters suggested by Ciesla et al. (2015).
The obtained photometric redshifts solutions are: z ∼ 0.4 (provided by prospector and LePHARE++), z ∼ 4 (very prominent solution for by eazy, also detected weakly by prospector and LePHARE++), z = 14−15 (obtained by all codes), z ∼ 17 (obtained by LePHARE++), and z ∼ 19 (obtained by bagpipes and LePHARE++).We note that the z ∼ 0.4 solution corresponds to a PAH line entering the F1000W passband, the z ∼ 4 solution would imply Paα covered by the filter, for z ∼ 15, Hα would contribute to the MIRI-F1000W flux, and z ∼ 19 means that the F1000W emission could be enhanced by Hβ and/or [O III]λλ4960,5008 emission.
We fitted the SED shown in Figure 4, which includes all flux data points with S/N > 3 and upper limits for the rest of bands, using the codes described above.We checked that our results concerning the photometric redshift and physical properties (for the two parameters mentioned in the main text, the stellar mass and the dust attenuation) do not change significantly if we only use S/N > 5 fluxes and upper limits, or just S/N > 5 fluxes.More specifically, the photometric redshift solutions are the same, but with different statistical weights (i.e., the peaks vary their relative strength).Concerning the physical properties, variations were very similar to the systematic differences between the distinct estimations for the two properties that are consistent among all codes and techniques, given that they are constrained by the two robust observational properties of Cerberus, i.e., the low luminosity and red color.More specifically, for the stellar mass we obtained variations of 0.3 dex when using higher signal-to-noise data for z ∼ 4 and z ∼ 15, and 0.7 dex for z ∼ 0.4.For the dust content, we obtained differences of < 0.3 mag in all cases, always remaining relatively high, above 3 mag, or finding a similarly obscured torus model.

Figure 1 .
Figure1.In the center of the figure, we show an RGB composition of the MIDIS F1000W field.The color images have been built with JADES data in 2 NIRCam filters, F277W and F356W, and the MIDIS MIRI F1000W filter (all images convolved to the same PSF of MIRI/F1000W).In the background, we show the HUDF JADES data in grayscale.We provide a series of zoomed-in RGB frames that lead to the NIRCam-dark MIRI source, Cerberus, whose image is located at the bottom right corner.The NIRCam-dark MIRI source Cerberus is highlighted with a square in that RGB frame.Additionally, we provide postage stamp images (2 ′′ ×2 ′′ ) of Cerberus in MIRI/F1000W, MIRI/F560W, and NIRCam/F444W (left column), showcasing that Cerberus (marked with an r=0.′′ 25 green circle) is clearly detected only at 10 µm.

Figure 2 .
Figure 2. Stacked NIRCam (left panel) and MIRI/F1000W (right panel) images of the region around Cerberus.The stacked NIRCam image was produced from individually PSF matched LW images.Note that the spurious pixels in the F460M image (see text for details) are removed in the median stacking procedure.The angular size of the images and zoomed inset is 3 ′′ ×3 ′′ and 0. ′′ 6 ×0.′′ 6, respectively.The red (blue) circle shows the location of the detected source in F1000W and has a radius of 0. ′′ 2. The green filled circle shows the resolution of the F1000W image (with a diameter of 0. ′′ 328).In the zoomed insets, we also show the centroid positions with the blue and red crosses, corresponding to the F1000W image and the NIRCam stack, respectively.

Figure 3 .
Figure 3. Photometric redshift probability distribution functions of Cerberus.All distributions are normalized to unity at the peak.We show the main solutions and also secondary solutions at z < 10 and z > 10.Different codes were used in the SED analysis, described in Appendix A.

Figure 4 .
Figure 4. Detailed analysis of the spectral energy distribution of Cerberus for the main redshift solutions.The symbols are the same as described in Figure 3.We provide best-fitting models provided by different codes that include both stellar and AGN emission.The top-left panel shows the best-fitting brown dwarf template from the ATMO2020++ models, for a Y-type dwarf with an effective temperature of ∼ 300K at a distance of ∼ 300 pc.The rest of the panels show results for z = 0.4 (top-right), z = 3.5 (bottom-left), and z = 15 (bottom-right).
data presented in this paper were obtained from the Mikulski Archive for Space Telescopes (MAST) at the Space Telescope Science Institute.The specific observations analyzed can be accessed via DOI: 10.17909/je9x-d314 DOI: 10.17909/je9x-d314.

Table 1 .
Photometric measurements for the Cerberus source.
Note-Thetable provides photometric measurements for Cerberus in all NIRCam individual images from JADES, incremental (named red and blue stacks) and disjoint (adding 2, 3, 4, and 5 contiguous filters) stacks, and MIRI bands from MIDIS.The central wavelengths and widths are given in µm, and the photometry in AB magnitudes, including 5σ upper limits for data points with S/N < 3.