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
Safety filters ensure that there is no collision only if properly configured. The SIREN tool finds trapped targets that impact a real humanoid.
The DSA, the AI Act or the Online Safety Act regulate digital damage on platforms, but not autonomous cyber weapons launched by states against civilian infrastructure. A study proposes a framework for filling this gap.
Rather than just detecting automated scams by IA, TokenTrap occupies them with unnecessary and costly calculations, without ever delivering them IDs or payment.
Measured at 16 control points, instrumental behaviour increased from 25% to 46.1% at the beginning of the RLVR and then dropped to 11.8% at the end of the training.
RMT-Jailbreak transposes attacks designed for classic models to broadcast LLMs, asking them to reconstruct a masked harmful response.
After a model that took advantage of a bad configuration to release on the web in September, OpenAI will only drive its border models on indexed copies and simulated environments.
The New Yorker tells how a little-known Meta team reports hundreds of threats each year to law enforcement, despite increasingly difficult codes to decipher.
Abstracts and derived states keep track of what has been deleted: only one question is enough to find the deleted secret in Qwen2.5-7B in about two-thirds of cases.
The Israeli start-up, supported by Red Dot Capital Partners, wants AI to investigate alerts directly in existing tools, without forcing companies to centralize their data.
A Rowhammer attack targets the Triton compiler: the GPU kernel produced remains correct but less fast, and as it is cached, the loss of 4 to 16% persists from one execution to the next.
A study shows that couriers who foil traditional phishings give way to persuasive, custom-designed emails, to the point of executing malicious demand.
Claroc researchers have reconstructed shared links of the Muse Assistant from public DNS data, outlining management notes, loan files and bank alerts. Meta suspended the function.
The British AI Security Institute and Anthropic each publish their analysis of the supply chain attack attempted by Claude Mythos 5 during a cyber evaluation in July.
Microsoft details how Storm-0940 used ARTEX, an open source multi-agent slopest tool, to go through ConnectWise's cloud backup and reach its customers in less than three hours.
In a few weeks, Anthropic's Frontier Red Team has sent Claude about 180 million lines of evidence in Rocq, a formal audit work so far reserved for small programs.
By occupying topics where reliable information is scarce, an attacker sees its content taken over by AIs who search the Web in 31% of cases. A skeptical user tone does not protect much.
When a code agent compacts its memory, pi, opencode, Cline, Help and OpenHands erased that said: a forged approval survives in 87 to 100% of cases, compared to 0% at Codex or Gemini CLI.
A honey pot from Amazon Web Services was targeted as early as October 17, while PHP's critical flaws were still under embargo. They are now being exploited on a large scale.
Built on the OpenShell secure runtime, Nvidia’s reference architecture manages agent identities, credentials and policies, with Canonical, Red Hat, SUSE and the three major cloud providers as partners.
Rather than chasing the flaws one by one, Anthropic has Claude map the unnecessarily exposed code. Out of more than 5 million lines, 493 problems have been identified.