AI Agents Don't Read Your Policy Docs. They Hit Your APIs — Gravitee
Démonstration en direct : une passerelle d'API bloque les appels destructeurs, l'extraction de données personnelles et les boucles d'un agent IA de réservation d'hôtel.
AM-IA brings together news, research papers and applications relating to artificial intelligence published around the world into a single feed — and makes them available in your European language at the click of a button.
Scroll to browse · Click to filter
Démonstration en direct : une passerelle d'API bloque les appels destructeurs, l'extraction de données personnelles et les boucles d'un agent IA de réservation d'hôtel.
A study compares four policies on the disclosure of open weights: if opponents find substitutes more quickly than the defenders, the restriction may backfire on the latter.
Researchers propose pre-calibrating the PO4AO controller using an internal light source, in order to correct for turbulence without the need for a linearly stable controller.
Out of 4,000 PNAS abstracts rewritten by 23 versions of LLMs, a detector that identifies 99 per cent of the rewrites before a generational change can identify only 3.8 per cent immediately afterwards.
Researchers have laid the foundations for a watch-smartphone system that anticipates the urge to smoke, based on interviews with ten smokers and former smokers in Asia.
A pre-recorded experiment involving 759 students shows that, after two preparatory sessions with a voice-based AI interlocutor, spontaneous participation in class increases significantly.
By combining reinforcement learning with the imitation of eight specialised policies, researchers are training a single multi-task controller, deployed on a Unitree B1 quadruped.
A test set of 10,000 false statements shows that four commercial generators reproduce them accurately more than seven times out of ten; an automated attack raises this rate to 95 per cent.
Hundreds of games and emulators, reconstructed using AI-assisted decompilation, have appeared in the space of a few weeks. Kotaku has tested web ports of classics that work almost perfectly.
The entire Deno team is moving to Cloudflare to make the Workers model usable outside its network. Deno Deploy will shut down in six months; Dahl highlights the AI agents.
Données privées, contenus non fiables, accès à Internet : Michael Patterson (Coder) montre pourquoi les agents de code installés sur les postes des développeurs cumulent les trois risques, et comment les cloisonner.
Back-cover blurbs, visuals, emails to literary agents: more than twenty employees tell Wired how AI is becoming established in the publishing industry, without the authors’ knowledge.
Launched by the CMS agency in August 2025, this forum for dialogue with the industry addresses the issue of AI applications’ access to medical records. Some legal experts see it as an advisory committee in all but name.
Jasmine Wang, Tomek Korbak and Mikita Balesni have written to the board of directors. OpenAI maintains that their dismissal relates to the handling of sensitive information, not to their security warnings.
It has the same architecture as Qwen-Image-2.1, with 7 billion parameters, but only eight denoising stages: the checkpoint is available on Hugging Face under a research licence.
In PE-Loop, the language model suggests revisions and features a deterministic calculation engine. Relying on an LLM for this assessment results in 58 per cent of non-compliant buildings being overlooked.
Guided by the geometry of buildings, a flow-matching model generates realistic 3D turbulent fields – with an error of less than 3 per cent in mean wind speed – to help cities adapt to climate change.
Healthcare professionals are primarily required to carry out administrative paperwork and search for information, with diagnosis accounting for just 3.7 per cent. The median benchmark does not include any documentation tasks.
Without deep learning, the method constructs a coordinate system based solely on the stellar spectra from APOGEE and DESI, and then applies the labels from a few stars to all the others.
There is no need for malice or hostility: according to two researchers, catastrophic risk arises from the interaction between misalignment, malicious use, organisational failures and the rush to roll out new systems.