Table of Contents

• 1) Introduction
• 2) Issues in krb5_child Helper
  • Upstream Fix
• 3) Issues in sssd_pam Helper
  • Upstream Fix
• 4) Notes on the selinux_child Helper
• 5) Notes on the ldap_child Helper
• 6) Issues in the sssd.service Unit
  • Upstream Fix
• 7) Further Observations
  • 7.a) Environment Variables
  • 7.b) Dumpable Process Attribute Setting
  • 7.c) Debugging Settings
• 8) Suggested Fixes
• 9) Situation on Other Distributions
• 10) Timeline
• 11) References

1) Introduction

SSSD (System Security Services Daemon) is a suite of daemons dealing with user authentication based on mechanisms like LDAP, Kerberos and FreeIPA. This report is based on SSSD release 2.10.0.

SSSD supports setting up user privilege separation by specifying the build time configure switch --with-sssd-user=.... The default in many Linux distributions is still to run SSSD as root, though. When privilege separation is enabled, then file based capabilities are assigned to a couple of helper binaries shipped by SSSD:

/usr/libexec/sssd/sssd_pam      root:sssd 0750 cap_dac_read_search=p
/usr/libexec/sssd/selinux_child root:sssd 0750 cap_chown,cap_dac_override,cap_setuid,cap_setgid=ep
/usr/libexec/sssd/krb5_child    root:sssd 0750 cap_chown,cap_dac_override,cap_setuid,cap_setgid=ep
/usr/libexec/sssd/ldap_child    root:sssd 0750 cap_chown,cap_dac_override,cap_setuid,cap_setgid=ep

Only members of the group of the dedicated sssd account are allowed to execute these privileged helpers. In SSSD before version 2.10.0 these helpers (with the exception of sssd_pam) had setuid-root bits. With commit 7239dd6791 this has been changed to using capabilities instead.

Our openSUSE SSSD packagers enabled privilege separation for the first time in conjunction with the update to version 2.10.0. This caused the privileged helpers to pop up on our radar, and we reviewed them. We found that these helper binaries do not currently provide proper privilege separation in SSSD. Some of them offer attack vectors to escalate to root again, or obtain powerful capabilities. Also the systemd service unit of SSSD has issues when privilege separation is active. The privileged helpers are not world-accessible, so no immediate exploitation by local users beyond the dedicated sssd account is possible.

We privately reported the findings described below to the Red Hat Security team on Nov 15. A coordinated disclosure process was in place for about a month, until the SSSD developers decided that the issues are not security issues, based on the following reasons:

• the issues are not directly exploitable, but only affect defense-in-depth.
• the sssd user and group are powerful by design, since these daemons influence the outcome of authentication.
• privilege separation has been introduced as additional hardening, not as a strong security layer.
• privilege separation was also introduced for some cosmetic purposes: “to allow running SSSD in restricted environment that do not support/allow running apps under uid=0/in user-ns (like restricted OCP profiles)”.

In our opinion, these issues are still security relevant. Consider for example a scenario where a system administrator or packager would allow execution of the privileged binaries to all users in the system, either by accident or by a false expectation of security. While it is common practice to deny world access to privileged binaries, in our experience this is usually done as a hardening measure only, not to protect against known weaknesses in programs. While the permissions of the helpers are correctly applied in SSSD’s installation routine, there is no documentation found that these helpers are security sensitive and must not be accessible to accounts other than sssd.

We did not press for CVE assignments, although we believe that formally they would still be justified. We recognize that it can still be an improvement to run SSSD processes as non-root, even if the privileged helpers allow escalation back to root.

Upstream has nonetheless worked and still works on a range of fixes to address findings from this report. The individual issues we identified are discussed in the following sections.

2) Issues in krb5_child Helper

Like most of the other helpers, this program accepts binary input on STDIN. The krb5_child helper reads the large and complex data structures struct krb5_req and struct pam_data. Interesting struct fields in this context are krb5_req.ccname and krb5_req.keytab, which specify file paths to process.

The ccname field is used in the code path privileged_krb5_setup() → k5c_ccache_setup() → k5c_precreate_ccache(). This ends up in a loop that creates all the parent directories of the path specified via STDIN, using uid and gid values also received from STDIN.

This proof-of-concept shows how to create arbitrary new directories with arbitrary ownership this way. This very likely allows a full local root exploit by skillfully creating directories under attacker control, e.g. directories used during lookup for trusted system binaries or libraries, or directories in /etc that are used for trusted configuration files or privileged services.

Upstream Fix

This specific escalation path has been addressed in the SSSD 2.10.1 bugfix release. Given the extensive interface offered by this helper program it is likely that further such escalation vectors exist. Since upstream does not consider this a strong security barrier, we have not looked any deeper into this component.

3) Issues in sssd_pam Helper

This helper starts a server instance which offers a socket-based IPC interface for passing PAM operation requests to it. This is the only helper that has more limited capabilities, namely only CAP_DAC_READ_SEARCH, which allows to override any read-access permission checks.

In the code path server_setup() → confdb_init() the following environment variables are interpreted in ldb_init() (which is part of Samba library code):

• LDB_MODULES_PATH
• LDB_MODULES_ENABLE_DEEPBIND
• TDB_NO_FSYNC

The LDB_MODULES_PATH variable allows the caller to specify a directory from which arbitrary shared objects are loaded via dlopen(). This simple proof-of-concept shows how to exploit this situation to gain access to the contents of /etc/shadow.

Contrary to the other helper programs, sssd_pam was not assigned setuid-root bits previously. The CAP_DAC_READ_SEARCH has only more recently been added via commit 0562646cc261, to allow it to access keytabs without having to run as root.

Upstream Fix

There is a pending upstream pull request to clear the environment of this helper to prevent this specific privilege escalation path.

4) Notes on the selinux_child Helper

We haven’t found any specific privilege escalation paths in this helper. The nature of this helper is to allow modification of SELinux MLS mappings, though. It accepts the new MLS range for arbitrary usernames via a binary protocol on STDIN. In an SELinux MLS managed system this is a pretty strong privilege for SSSD to have. Since this is likely by design, we suppose there is little that can be done about that, except for documenting the sensitive nature of the helper and that access to it must be well restricted.

The helper performs calls into shared SSSD library code and into libsemanage and libselinux. Luckily we couldn’t find any cases where overly problematic environment variables are interpreted (beyond the common variables listed in section 7.a).

This helper also changes its UID and GID to 0 early on. When transitioning to UID 0, the kernel does not assign the full set of capabilities to the process again. This means that the process runs under restricted root privileges, having UID and GID 0 but only the capabilities assigned to the selinux_child binary. Additionally, the SSSD processes run with the systemd hardening feature SecureBits=noroot noroot-locked, thus preventing the helper from using setuid binaries like sudo to regain full root privileges.

The security of restricted root privileges in Linux is lacking, though. The Linux kernel uses capabilities for its permission checks, but userspace utilities normally only rely on other process’s UID and GID credentials. Also some APIs lack the possibility to express restricted root privileges, e.g. in UNIX domain sockets the SO_PEERCRED option is used to determine the credentials of a peer process, which only provides a struct ucred, containing the peer process’s PID, UID and GID. Thus this helper runs with privileges close to full root.

5) Notes on the ldap_child Helper

We couldn’t find any bigger problems in this helper, beyond the generic comments in section 7) that also apply to this helper.

6) Issues in the sssd.service Unit

The sssd.service systemd unit contains the following ExecStartPre lines:

ExecStartPre=+-/bin/chown -f -R root:@SSSD_USER@ @sssdconfdir@
ExecStartPre=+-/bin/chmod -f -R g+r @sssdconfdir@
ExecStartPre=+-/bin/sh -c "/bin/chown -f @SSSD_USER@:@SSSD_USER@ @dbpath@/*.ldb"
ExecStartPre=+-/bin/chown -f -R @SSSD_USER@:@SSSD_USER@ @gpocachepath@
ExecStartPre=+-/bin/sh -c "/bin/chown -f @SSSD_USER@:@SSSD_USER@ @logpath@/*.log"

The directories /var/log/sssd and /var/lib/sssd are owned by the unprivileged sssd user. The chown and chmod lines above, which are run as root, allow a compromised sssd user to stage symlink attacks and thus gain ownership of, or access to, privileged system files.

This is a simple proof-of-concept demonstrating the issue:

# stage a symlink attack to gain ownership of /etc/shadow
sssd$ cd /var/log/sssd
sssd$ ln -s /etc/shadow my.log
# as root trigger a sssd (re)start
# sssd needs to be configured (i.e. /etc/sssd & friends need to exist) for this to work
root# systemctl restart sssd.service
root# ls -lh /etc/shadow
-rw------- 1 sssd sssd 889 Nov 13 11:40 /etc/shadow

As the directories that are affected by this are not world-writable and don’t carry a sticky bit, the Linux kernel’s symlink protection does not come to the rescue here. Path arguments that are named directly on the command line of chown or chmod will be followed, if they’re symlinks, unless --no-dereference is passed. Passing this option is also the recommended fix for this.

In our opinion such automatic permission “fixes” should be treated with care. If this is to avoid any trouble with migration from older installations (without privilege separation) then we would rather offer an explicit utility for system administrators to run. This would make clear that the logic only runs once and not every time sssd is started. It also would prevent any configuration errors from persisting or from being masked (e.g. other components in the system that assign bad permissions to SSSD files, thus fighting against the automatic permission fixes).

Upstream Fix

There is a pending upstream pull request to pass --no-dreference to the chown invocations found in the systemd service unit.

7) Further Observations

7.a) Environment Variables

There are further environment variables interpreted by the helper programs:

• TALLOC_FREE_FILL: will cause memory free()‘d via talloc_free() to be overwritten with the byte set in this variable.
• _SSS_DOM: will influence systemd journal log messages and thus allow a bit of log spoofing.

While these variables have only minor influence on program execution, privileged programs should not allow arbitrary environment settings to affect their behaviour.

7.b) Dumpable Process Attribute Setting

All of the privileged helpers support a --dumpable command line switch to control whether the process will have the dumpable bit set or not. The default for this even is to mark the process as dumpable (SUID_DUMP_USER). This somewhat unexpectedly overrides the sysctl setting fs.suid_dumpable, which is usually 0 or 2.

The dumpable setting of a process is a sensitive property that plays an important role in the ptrace() system call to determine whether tracing another process is allowed. From man 2 ptrace:

These checks are performed in cases where one process can inspect sensitive information about, or in some cases modify the state of, another process. The checks are based on factors such as the credentials and capabilities of the two processes, whether or not the “target” process is dumpable […]

We believe the only barrier left that prevents the unprivileged sssd user from being allowed to trace the privileged binaries is this (further excerpt from man 2 ptrace):

(5.2) Deny access if neither of the following is true: - The caller and the target process are in the same user namespace, and the caller’s capabilities are a superset of the target process’s permitted capabilities.

Attaching via ptrace() is only denied because the target processes have raised capabilities. However, this is only a kind of kernel security extension, provided by the kernel security module security_ptrace_access_check().

Besides this, the dumpable setting allows the unprivileged user to send e.g. a SIGSEGV signal to the privileged processes and force them to dump core. What happens from here depends on the core dump handler installed in the system. systemd-coredump safely handles such core dumps and the unprivileged user cannot access them. If only core is configured as a core pattern, like it is the case on Debian Linux by default, for example, then the unprivileged user can cause the core file to be created in arbitrary directories, by first changing into them, starting the privileged process, then killing it. The core file will not be readable for the unprivileged user, but it still allows to clutter the file system and maybe even overwrite legit files that are named core.

7.c) Debugging Settings

The privileged programs also offer rich command line settings for enabling debugging output and redirecting it to various locations. Generally, privileged programs should be very careful about what kind of information is leaked to the caller. The debugging logs can contain information that weakens security features (like stack overflow protection) or leak privileged information that has been read in from privileged files.

8) Suggested Fixes

For privileged setuid-root-like binaries the usual precautions should be taken:

• change into a safe current working directory (CWD).
• apply a safe umask (this already happens)
• cleanse the environment from any untrusted variables, only keep a whitelist of vetted variables. E.g. also set a safe PATH.
• make sure none of the interfaces (command line parameters, STDIN data input) offers possibilities for the unprivileged caller to escalate their privileges beyond the scope of what the privileged program is supposed to do.

The last item will likely be the most difficult to realize for the programs in question - especially in the krb5_child helper we expect more attack surface to exist, for example in the handling of the ccache and keytab files. These are dealt with in various other code paths via krb5 library routines that are unaware of the untrusted input. We suggested upstream to carefully think through all the possible inputs and code paths and to tighten them.

We would furthermore strip down the supported command line switches, or limit critical switches to callers that are root, notably the switches that influence debugging and logging as mentioned in section 7.c.

The dumpable setting should be left unchanged in the privilege escalation context.

Finally we suggest to clearly document what can be expected of the privilege separation feature and how the privileged helpers need to be packaged in order to achieve a safe installation (especially that they must not be world executable). This already happened in the description of the 2.10.1 bugfix release.

9) Situation on Other Distributions

We looked into a number of other Linux distributions and found that on Fedora, Debian and Ubuntu the SSSD privilege separation is not currently used. On Arch Linux the current 2.10.0 version is used together with privilege separation, though, which is affected by the issues covered in this report. Upstream informed us that there are plans for Fedora Linux to use the privilege separation feature soon as well.

10) Timeline

11) References

• sssd GitHub project
• sssd 2.10.1 release containing a first batch of fixes addressing issues from this report
• Pending hardening pull request containing further fixes

Hace mucho tiempo publiqué el mapa generado en Google de los usuarios de KDE. Ese mapa era colaborativo y voluntario pero generado sobre Software Privativo. En varias ocasiones pedí que alguien lo cambiara a OpenStreetMap y mis súplicas fueron escuchadas. Ya casi había olvidado este proyecto pero este agosto lo recuperé, publicitándolo en el grupo de Telegram KDE – Cañas y Bravas y parece que el proyecto ha tomado impulso, así que he decidido seguir ppublicitándolo durante unos meses. Así que, tras un mes de descando, os doy la bienvenida a la actualización de diciembre_24 del mapa de usuarios KDE que, como vais a ver, se ha sigue llenand pero todavía no tenemos a nadie en Cáceres ,Lleida, Cuenca, Salamanca o Huesca, por poner algunas capitales de provincia. ¿No podemos solucionar este vacío?

Actualización de diciembre_24 del mapa de usuarios KDE

Si vais a ver el mapa de la Comunidad KDE de España marzo 2018, podréis ver una imagen estática de los usuarios que se habían puesto voluntariamente , y digo estática, porque alguien borró (queriendo o sin querer) los «pines» del mapa.

Ese incidente fue comentado en el grupo de Telegram KDE – Cañas y Bravas, y un usuario dijo que trabajaría en ello aprovechando que se debía empezar de nuevo el mapa.

Este compañero de grupo era @wakutiteo y creó, utilizando los servicios de uMap sobre OpenStreetMap el nuevo mapa de la usuarios KDE en España en el que podemos empezar a registrarnos.

En un principio había poca poca gente apuntada (mirad la entrada de junio de 2018) pero poco a poco se ha ido llenando, a pesar de la poca promoción que ha tenido. A ver si a partir de ahora voy dándole más cancha y el mapa crece . De momento pongo la captura de pantalla de este mes de agosto, la del mes de septiembre y la de este octubre del 2024 para que veamos la evolución, lenta pero constante.

Agosto 2024

Septiembre 2024

Octubre 2024

Diciembre 2024

Como vemos, siguen apareciendo poco a poco más usuarios, con una espectacular crecimiento en Andalucía, Galicia, todo el norte y este peninsular.

Pero seguimos sin usuarios en capitales de provincia como Cáceres, Lleida, Cuenca, Salamanca, Huelva, Ciudad Real, Albaceta, Zamora, Lugo, Soria, Teruel, Segovia, Tarragona, Girona, Alicante, Guadalajara, Santiago de Compostela, Pontevedra o Ávila, si no he mirado mal (que es posible).

Evidentemente, es voluntario y apuntaros es bajo vuestra propia responsabilidad. Os dejo también el mapa incrustado que se irá actualizando poco a poco.

Mostra-ho a pantalla completa

Pincha para ver en pantalla completa

Y también os dejo un vídeo de cómo editar el mapa para poder apuntaros.

¿Qué es Umap?

Básicamente, con uMap puedes crear mapas con capas de OpenStreetMap en un minuto e incrustarlo en tu página web, de código libre. Con uMap puedes realizar las siguientes acciones:

• Elegir las capas de tu mapa
• Añadir PDIs: marcadores, líneas, polígonos…
• Elegir los colores y los iconos de los PDIs
• Gestionar opciones del mapa: mostrar un minimapa, localizar al usuario al cargar…
• Importar por lotes datos geoestructurados (geojson, gpx, kml, osm…)
• Elegir la licencia de tus datos
• Embeber y compartir tu mapa

La entrada Actualización de diciembre_24 del mapa de usuarios KDE se publicó primero en KDE Blog.

Esta semana compramos muita pipoca na MultiCortex, pois foi um absoluto cinema de papers no setor de IA. Acreditamos que foi a semana mais insana de 2024 marcada por avanços significativos no campo da inteligência artificial, com a publicação de diversos artigos que abordam desde melhorias em arquiteturas de modelos até questões de segurança e eficiência. A seguir, a lista desses trabalhos:

• Byte Latent Transformer
• Training Large Language Models to Reason in a Continuous Latent Space
• Language Modeling in a Sentence Representation Space
• Phi-4 Technical Report – Best-of-N Jailbreaking
• Forking Paths in Neural Text Generation
• Refusal Tokens – [MASK] is All You Need
• Explore Theory-of-Mind
• Obfuscated Activations Bypass LLM Latent-Space Defenses
• The Pitfalls of Memorization
• How to Merge Your Multimodal Models Over Time?
• Machine Unlearning Doesn’t Do What You Think
• Understanding Gradient Descent through the Training Jacobian
• An Evolved Universal Transformer Memory
• Transformers Struggle to Learn to Search
• Transformers Can Navigate Mazes With Multi-Step Prediction
• Frontier Models are Capable of In-context Scheming
• Mixture of Monosemantic Experts for Transformers
• Learning to Merge LoRAs via Hypernetworks for Subject-Style Conditioned Image Generation
• Scalable Text and Image Conditioned Video Generation
• Hidden in the Noise: Two-Stage Robust Watermarking for Images
• Learned Compression for Compressed Learning
• Learning Flow Fields in Attention for Controllable Person Image Generation
• ProcessBench: Identifying Process Errors in Mathematical Reasoning
• Unraveling the Complexity of Memory in RL Agents
• Expanding Performance Boundaries of Open-Source Multimodal Models with Model, Data, and Test-Time Scaling
• APOLLO: SGD-like Memory, AdamW-level Performance
• Neural LightRig

The openSUSE community will celebrate the 5-year anniversary of the openSUSE Bar on Dec. 19. Join people in the bar and celebrate this social space where open-source enthusiasts, developers and like minded individuals can come together to discuss open source and other topics.

The bar has become more than just a meeting place; it is a space to collaborate, connect and a welcoming atmosphere for sharing ideas. For years, the space has had countless informal meetups and networking events, celebratory moments after openSUSE milestones and is a place for openSUSE fans to gather.

People are invited to join in the celebration and to share memories or stories of their experience in the openSUSE Bar.

You can find a video about the beginning of the openSUSE Bar from a talk at openSUSE Conference 2022.

La web HTML para personas ahora también está disponible en formatos Markdown y PDF

Hace ya una semana que publiqué por el blog el proyecto en el que estaba enfrascado, que no es otro que HTML para personas.

• https://victorhck.gitlab.io/htmlparapersonas/

La acogida ha sido genial, muchas gracias a todas las personas que me han hecho llegar su opinión y críticas, que siempre han sido buenas y positivas, me alegro.

Hoy quiero anunciar una novedad. El texto de la web está disponible en formato Markdown y en PDF.

Durante esta semana he estado trabajando en hacer posible que la web esté disponible también en estos dos tipos de formatos. Ha sido un proceso «casi» artesanal de revisión y comprobación. Espero que todo funcione como debe.

Y ya está disponible HTML para personas:

• en formato Markdown
• y en formato PDF

Están los enlaces también en el pie de la propia página para acceder desde ella.

Todo el código de la página, imágenes y estos dos documentos están disponibles desde su repositorio GitLab.

Y nada más que añadir…

CentOS Stream 10 and EPEL 10 just became available, and as usual, I tried to build syslog-ng as soon as possible. For now it is available in my git snapshot repository, but I am also planning to make it available in EPEL 10 soon.

Read more at https://www.syslog-ng.com/community/b/blog/posts/test-syslog-ng-on-epel-10

Tras un parón debido al salto de Qt5/KF5 a Qt6/KF6 que realizó la Comunidad KDE hace ya más de seis meses decidí retomar esta sección aunque renombrándola ya que en ella solo hablaré de Plasmoides para Plasma 6. Así que hoy os presento el lanzador de aplicaciones Orbit Menu, el plasmoide número 13 de esta nueva serie.

Lanzador de aplicaciones Orbit Menu – Plasmoides para Plasma 6 (13)

Como he comentado en otras ocasiones, de plasmoides tenemos de todo tipo funcionales, de configuración, de comportamiento, de decoración o, como no podía ser de otra forma, de información sobre nuestro sistema como puede ser el uso de disco duro, o de memoria RAM, la temperatura o la carga de uso de nuestras CPUs.

Así que espero que le deis la bienvenida a un nuevo lanzador de aplicaciones Orbit Menu una creación del incombustible zayronXIO que nos ofrece otra variante de este tipo de plasmoides que destaca por la posición central del avatar del usuario y la barra inferior de acciones del sistema, algo poco habitual en este tipo de widgets.

Y como siempre digo, si os gusta el plasmoide podéis «pagarlo» de muchas formas en la página de KDE Store, que estoy seguro que el desarrollador lo agradecerá: puntúale positivamente, hazle un comentario en la página o realiza una donación. Ayudar al desarrollo del Software Libre también se hace simplemente dando las gracias, ayuda mucho más de lo que os podéis imaginar, recordad la campaña I love Free Software Day de la Free Software Foundation donde se nos recordaba esta forma tan sencilla de colaborar con el gran proyecto del Software Libre y que en el blog dedicamos un artículo.

Más información: KDE Store

¿Qué son los plasmoides?

Para los no iniciados en el blog, quizás la palabra plasmoide le suene un poco rara pero no es mas que el nombre que reciben los widgets para el escritorio Plasma de KDE.

En otras palabras, los plasmoides no son más que pequeñas aplicaciones que puestas sobre el escritorio o sobre una de las barras de tareas del mismo aumentan las funcionalidades del mismo o simplemente lo decoran.

La entrada Lanzador de aplicaciones Orbit Menu – Plasmoides para Plasma 6 (13) se publicó primero en KDE Blog.

O laboratório da MultiCortex começa testar os modelos Apollo da Meta que estão disponíveis em três tamanhos (1,5B, 3B e 7B) e processam até uma hora de vídeo. Eles se destacam na compreensão avançada de vídeos, respondendo a perguntas complexas, interpretando a linha temporal e sustentando diálogos sobre o conteúdo exibido. O Apollo-3B rivaliza ou supera modelos de 7B, enquanto o Apollo-7B supera até modelos de 30B+ em benchmarks de vídeo, demonstrando eficiência superior e excelente relação custo-benefício.

O Apollo LLM é uma nova geração de modelos de linguagem multimodal (LMMs) que redefine a forma como as máquinas entendem vídeos. Fruto de uma colaboração entre a Meta GenAI e a Universidade de Stanford, o Apollo se destaca por sua eficiência e inovação no campo da inteligência artificial.

Principais Diferenciais

1. Consistência de Escalonamento (“Scaling Consistency”)
   O conceito de “Scaling Consistency” permite que decisões de design feitas em modelos menores sejam aplicadas de forma confiável a modelos maiores. Isso reduz drasticamente o custo computacional e o tempo de desenvolvimento.
2. Otimização de Estratégias de Treinamento
   O Apollo explora cronogramas de treinamento e combinações de dados. A descoberta de que a amostragem “fps” (frames por segundo) é mais eficiente que a amostragem uniforme impacta diretamente o desempenho do modelo.
3. Tokenização Eficiente
   O modelo encontrou o equilíbrio ideal de 8 a 32 tokens por frame, além de usar técnicas de “token integration” para integrar tokens de diferentes quadros de maneira eficaz.
4. Benchmark Personalizado (ApolloBench)
   Para medir a eficácia dos LMMs de vídeo, o Apollo introduziu o ApolloBench, um benchmark personalizado que facilita a avaliação de diferentes modelos.
5. Uso de Encoders de Alto Desempenho
   O SigLIP-SO400M foi identificado como o melhor encoder único, mas a combinação deste com o InternVideo2 proporcionou um desempenho geral ainda mais robusto.
6. Aprimoramento Contínuo
   O modelo melhora a performance de seus encoders em dados de vídeo e ajusta a mistura de dados de texto e vídeo, garantindo desempenho superior em tarefas de raciocínio e percepção.

Aplicações Práticas

O Apollo LLM se destaca em tarefas que exigem entendimento multimodal, incluindo:

• Assistentes de IA mais inteligentes
• Análise de vídeos para segurança e vigilância
• Interpretação de vídeos para ensino e aprendizado

Desempenho Comparado

O Apollo superou diversos concorrentes em benchmarks de raciocínio espacial, percepção e compreensão de conteúdo de vídeo. Seu desempenho foi validado em testes no TempCompass, MLVU e VideoMME.

O Apollo LLM é uma solução de ponta que promete transformar a interação homem-máquina em ambientes multimodais. Para mais detalhes, acesse o site oficial do Apollo.

https://arxiv.org/pdf/2412.10360