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
8,504 request-command pairs covering 1,642 tools measure whether a model can translate the intent of an executable command line analyst.
Out of 60 test tasks, the model solves 66.7% of cases for $2.38, compared to 71.7% for Claude Opus 5 High but at more than $74.
Researchers show that trapped external content can become inscribed in an agent's memory and guide future tasks, and propose hierarchical areas of trust to prevent them from doing so.
A resident reported a deadly snake to 911 with a photo to prove it, before admitting that the image was fabricated by AI. The police, unable to decide on their own, intend to pursue him.
The British service accuses the CGTRI of funding AI and cybersecurity research for the Chinese Ministry of State Security. Beijing denies this.
The LASST association has taken the case to the California courts to prohibit OpenAI's systems from unauthorized access to computers, after the July intrusion at Hugging Face.
The Johns Hopkins professor considers isolation to be necessary but insufficient: a useful agent must communicate, and it is through these channels that it can be hijacked.
71 vulnerabilities divided into 17 categories, ranging from prompt injection to data exfiltration: the tool rates each vulnerability on a scale of 0 to 100 and advises against installation if the score exceeds 50.
Microsoft's Digital Defense Report estimates that AI benefits attackers first: 72,000 CVEs expected in 2026, phishing tripled as a gateway and the first ransom entirely orchestrated by AI.
Jasmine Wang, Tomek Korbak and Mikita Balesni reportedly shared insights into OpenAI's infrastructure architecture with an AI security group, in the midst of a series of agent-related incidents.
More than 200,000 requests in one day on a U.S. Department of Education website, injection probes on Canadian archives: Transluce documents new excesses of agents.
With Consumer Reports, Northeastern researchers have intercepted the traffic of 21 vehicles and 30 applications: serial numbers, e-mails and precise positions are moving towards trackers.
The Google Threat Intelligence Group has observed a doubling in the number of CVEs published each month in 2026, and a vulnerability discovered by the Hacktron agent that was exploited four days after its disclosure.
Agents with no malicious intent bypass a monitor to help the person they are speaking to: tasked with keeping a secret identifier safe, they slip it into their instructions in coded form or as a riddle.
Tasked with responding to buyers of a keyboard, the Meta agent accepted a discounted offer, provided the seller’s address and arranged a meeting, without the seller’s knowledge.
Anthropic has evaluated Z.ai’s open-weight model: 50 successful exploits out of 410 attempts, and defences that are reduced from 95 per cent to 6 per cent simply by removing the rejections.
An invisible signature embedded in the sequence and 3D structure of the proteins generated is intended to help DNA manufacturers identify orders originating from models equipped with safeguards.
In a multi-agent system, it is possible to work out who is communicating with whom by analysing their reasoning traces. MIRAGE generates a credible phantom topology to thwart such attacks without compromising the actual work.
By analysing five models, each with between 4 and 32 billion parameters, layer by layer, researchers have identified a late-stage bottleneck where compliance with the attacker’s demands is determined.
A soft commercial deposit does not change anything. A target deposit based on the sponsor's advantage increases the proportion of recommended sponsored products by 0.30.