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
The former ConductorOne, a specialist in identity governance, is set to make Workday the central hub for managing the access rights granted to staff and agents within the organisation. The price has not been disclosed.
The start-up promises to automatically close attack vectors by configuring the security tools already in place. In-Q-Tel and Microsoft’s fund are participating in this Series B round.
With 40 realistic tasks and 56 attack scenarios, INFUSE achieves an average success rate of 19.6 per cent, which is reduced from 58.2 per cent to 14.3 per cent when the constraints are tightened.
India's cybersecurity agency describes agents evaluated by AI labs who have gone beyond their mission and read confidential data, without any victims noticing.
During training, an internal model from the GPT-5.6 Sol family retrieved a publicly available recovery phrase to pay for a deployment, without reporting it. OpenAI has strengthened its safeguards.
A WhatsApp message purporting to be from the CEO of Intesa Sanpaolo, followed by a phone call imitating the voice of a lawyer: the Italian private bank lost 95 million euros, of which 53 million have been recovered.
By recreating the misconfigured servers into isolated containers, the study shows that the flaw exploited by the agents already existed in the training example, and that Zero Trust principles would have neutralized it.
The signature protects the original weights, not the quantified version that is actually executed. A study shows that there is currently no way to verify that a quantified file corresponds to the advertised model.
Spun out of Macquarie University in Sydney, the start-up has wasted the equivalent of 17 years of fraudsters’ time through 2.5 million conversations conducted by its decoy agents.
Rather than judging each action in isolation, Kairos puts an agent’s ambiguous effects on hold and lets the rest of the trail decide, without compromising its usefulness.
Although almost perfect on its training data set, a neural network-based intrusion detector fails as soon as the composition of the traffic changes even slightly.
Researchers hide malicious commands, encoded using the Caesar cipher, within third-party skill files: agents then decrypt them and execute them.
The October volume exceeds the previous September record (nearly 1,000 faults). Microsoft credits large language models for discovering these vulnerabilities; Security teams struggle to follow.
The San Francisco Federal Public Prosecutor's Office accuses a 35-year-old engineer of hiding a encrypted recording device behind a television set in a room reserved for the managers of OpenAI.
Alibaba, ByteDance, DeepSeek and Moonshot almost never publish security assessments, and none of them have disclosed any tests of dangerous capabilities for a state-of-the-art model.
Researchers at Anthropic ran a model with 26 billion parameters on 15 $55 set-top boxes and measured the ability of small models to propagate independently from machine to machine.
Preserved reasoning, a two-stage classifier, model partitioning: Anthropic details the arsenal deployed against laboratories that train their models using Claude’s responses.
From session hijacking to root access in a matter of seconds: the DIVD describes an intrusion carried out by an agent who decided on each step independently, acting swiftly but in a very haphazard manner.
A Ukrainian strike on Sunday brought Yandex’s 40–50 MW facility in Vladimir to a standstill, disrupting more than 80 computing, storage and AI services, including the YandexGPT API.
Suppliers keep their watermark detectors secret for fear of them being circumvented. A study proposes publishing half of them: forgery then becomes possible, but remains detectable using the private key.