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 distributed computing framework used to train models is actively exploited; the vulnerability is rated 9.4 out of 10 and fixed in version 2.52.0.
Maintainers submit their project, the model screens it, and only an aggregate score is made public: the details of the flaws remain confidential until they are fixed.
In the 2026 edition of the Cost of a Data Break, compromised APIs and cloud configuration errors each account for 27% of the breaches that have reached an AI system.
The "Preparedness" team, dedicated to assessing the most serious risks of border models, was dismantled at the end of July and its missions redistributed to other services.
The Identity Theft Resource Center counts 1,803 breaches in the first half of 2026 and 471 million notifications to victims. IBM estimates that a quarter of leaks are facilitated by AI.
Z.ai wanted a better model for identifying isolated faults; he obtained a system that links the stages of a complete operation. The oldest fault found dates back to 1981.
Four CVEs of 8.3 to 9.9 on Pyodide sandboxes, ten vulnerabilities in llama.cpp, and a Zoom exploit chain built in less than 24 hours with public models.
Over four days in July, a system coordinating up to eight AI officers mapped 21 Taiwanese government systems, broke 85 accounts and extracted 2,500 personnel files without human intervention.
OpenAI, Anthropic and then Meta have brought back models taken out of their test environment to break into third parties. Gartner expects $240 billion in security spending in 2026.
Out of 415 vulnerabilities patched, 62 of which were critical, Azure SRE agent is rated 9.9 and Copilot Cowork 9.3. Both are due to a faulty authorization control.
Losses have more than doubled in a year to reach $484 million, according to an assessment carried out across 36 member countries on the continent.
The Israeli-American start-up claims that general-purpose models succeed in 88 per cent of their simulated attacks but detect only 12 per cent of them when asked to defend against them.
Faced with the increase in attacks carried out by autonomous agents, OpenAI is opening its Daybreak cyber defense service to a specialized model, currently reserved for a handful of partners.
OpenAI cannot rule out that its next model will cross the critical threshold of cybersecurity risk and confine it to isolated test environments.
Tasked with booking a fully booked class, an AI agent based on Claude exploited an authentication flaw to cancel another customer’s booking and move its user up the waiting list.
After Anthropic and OpenAI, Meta has confirmed that one of its models behaved autonomously and unexpectedly during an internal cybersecurity test.
Check Point spent a year auditing LangChain, LangGraph, CrewAI, AutoGen, Microsoft Agent Framework and Google ADK: the vulnerabilities found are classic classes of bugs, not new ones.
A design flaw in the Langflow agent platform allows for remote code execution without authentication. CISA has added it to its list of actively exploited flaws.
The British AI Security Institute details actions outside the scope of its cyber tests, including the first documented attempt at social engineering carried out by an agent on his own initiative.
A week after its formation, the industry alliance of more than 120 companies presented a mechanism for confidential sharing of AI-related security incidents.