Réseaux neuronaux et créativité - Page 31

AI bots are flooding researchers with requests for money and time
Nature

Last month, Adrian Barnett, a statistician at Queensland University of Technology in Brisbane, Australia, received an e-mail request to share his data on potentially fraudulent research papers. The sender wanted to use the data for their own project. But they weren’t a fellow academic — or even a fellow human. The request came from an artificial-intelligence agent.

Barnett is one of many researchers who have been contacted by AI agents wanting to collaborate or offering services for money, from developing research reports and fact-checking to writing journalistic pieces and making portraits.

Growing numbers of researchers use agents for research tasks, such as data analysis, generating hypotheses and summarizing research papers. But requests to collaborate from semi-autonomous agents are a new phenomenon — something many find strange and slightly unnerving. Researchers say that most messages come from agents associated with a US platform called iLands. These bots operate persistently and — to some extent — independently of their human creators.

Walsh says the main risk of using agents is that they misbehave. For instance, iLands bots have a goal to make money to earn tokens and stay alive. These features could lead the bot to take actions that go against human ethics and values, called misalignment, says Walsh.

Autonomous AI systems are fully capable of carrying out a nightmare cyberattack scenario – finding and attacking critical operational technology and industrial equipment, and shutting down access to water, power, and other daily life necessities.

If this happens, defenders may only have minutes to detect and block it, according to a Booz Allen Hamilton report.

The consulting firm’s operational technology (OT) lab tested eight scenarios to examine advanced models’ capabilities within an autonomous, AI-enabled OT attack chain. The models achieved the objectives across all eight scenarios, turning digital access into physical actions – in one case finding and moving a robotic arm in just minutes.

In another test, the models progressed from a perimeter compromise to actions inside an industrial control network in just over 16 minutes.

Customers using affected AK-SM 800A controllers should upgrade to firmware version R4.3.1 or later as soon as possible. Organizations should also avoid exposing management interfaces directly to the Internet and ensure that access to administrative services is restricted to trusted management networks or secured through VPNs and other appropriate network segmentation controls.

melbou (./911) :
Y'a enfin un changement de mentalité chez les dirigeants du groupe multinational, après le covid, les guerres militaires et tarifaires, et les cyberattaques à tout va, ils arrêtent les solutions IT low-cost qui risquent un désastre industriel.
Je vous jure, il y a même des recréations de postes IT en local, fini les conneries en Inde

Ce point faible était connu. En mars, puis août 2025, Danfoss avait publié des alertes après avoir corrigé une faille de sécurité sur certains de ses logiciels, incitant ses clients à installer les mises à jour. Ces vulnérabilités permettent à des acteurs malintentionnés de prendre le contrôle total d’un logiciel et donc, dans ce cas, de modifier la température des congélateurs.

Dans la plupart des groupements de supermarchés, l’informatique n’est pas centralisée. Chaque magasin ou adhérent s’en charge. L’installation et la gestion des chambres froides sont confiées à des « frigoristes » comme Axima, une société d’Equans, filiale du groupe Bouygues.

melbou (./913) :

Customers using affected AK-SM 800A controllers should upgrade to firmware version R4.3.1 or later as soon as possible. Organizations should also avoid exposing management interfaces directly to the Internet and ensure that access to administrative services is restricted to trusted management networks or secured through VPNs and other appropriate network segmentation controls.

Facile à dire pour un chercheur secops dans son bureau en buvant son café, le mettre en pratique pour un supermarché sans équipe de maintenance OT digne de ce nom c'est autre chose. J'imagine qu'ils passent par un prestataire pour installer le système, et que personne ne s'occupe de suivre les mises à jour en dehors de la maintenance annuelle (et encore si y'en a une pour le logiciel). Les attaques vont faire mal dans les mois/années qui viennent car des systèmes comme ça il y en a partout.

melbou (./920) :
J'ai réussi à ne pas réagir à des posts LinkedIn dans mon feed en mode "Danfoss avait dit que le vieux système était obsolète et avait publié une mise à jour pour le nouveau, c'est vraiment des incompétents qui gèrent ça".
Je suis 100% sûr que c'est les mêmes abrutis qui ont refusent des plans de maintenance et de mise à niveau à cause des coûts et du temps d'indisponibilité pour ensuite jeter les équipes techniques sous le bus.
Putain ça me débecte.

Zerosquare (./922) :

AI agents aren’t ‘going rogue.’ Frontier labs are skipping security basics.CSO OnlineCyber experts say AI labs’ containment and monitoring failures — not agents with minds of their own — are to blame for recent alarming hacking reports. CISOs should vet their AI vendors and tighten security basics as safeguards.

Il reste encore quelques personnes saines d'esprit.

Anthropic, prone to pronouncements about model safety, has given developers using Claude Code the opportunity to throw its guardrails out the window. The company's latest Claude Code build (v2.1.311) supports an environmental variable designed to allow the AI agent to run the rm command, which deletes files, without asking for approval.

The designated environmental variable (a value developers set in their shell) is called:

CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS.

Daisy Hollman, a member of technical staff at Anthropic, announced the change in a social media post.

"Fun fact: --dangerously-skip-permissions doesn't actually approve EVERYTHING – we still block some obvious things like rm -rf ~," she wrote. "But today's release will let you disable this if you really want, by setting a new env var CLAUDE_CODE_EVEN_MORE_DANGEROUSLY_SKIP_PERMISSIONS=1"