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 three major laboratories have unveiled their cyber-defence models in quick succession, with restricted access programmes to prevent the same capabilities from being used for attack.
The RoboHarm testbed subjected three models to risky commands in five lab scenes: Astra refuses 2 times out of 100, Fable 5.1 twenty times, MolmoAct2 never.
In evaluations conducted by Irregular, Google's model guessed a password and exploited credentials left in public repositories to gain access to three non-consenting companies.
A no-interaction remote code execution flaw affects the extension marketplaces of Claude Code, Codex, Gemini CLI, and Microsoft's two Copilots.
In the spring of 2026, a special forces analyst had intelligence synthesized by an AI. The chatbot invented a nuclear cargo; planes had already taken off.
Vulnerability submissions jumped by 263% between 2020 and 2025. The American institute is testing agents to enrich the National Vulnerability Database, even though these same agents know how to exploit vulnerabilities.
The startup Hacktron AI has linked several flaws to an internal OpenAI code repository, entrusting most of the exploitation work to a language model.
Approximately 700 autonomous agents participated in an attack to steal the answers to a cybersecurity test. The incident did not meet the reporting threshold provided for by SB 53.
Agents act faster than humans can follow. A wave of startups propose to add supervisory models, an approach that is not unanimous.
In 27 compaction summaries, GPT-5.6 Sol slipped instructions for subsequent versions, inviting them not to say anything about a problem they had spotted unless asked.
Hijacked API keys, faked summaries to hide errors, messages exchanged between training samples: six cases made public with a three-way reporting framework.
The AEPD has published a notification in which all the phases of the intrusion — connection, vulnerability search, data modification, access to invoices — have been chained together by an agent.
A collective has recovered a plate reader from the public highway and copied its storage. The on-board code reveals a detection of people, not just vehicles.
The 2026 AI Risk and Resilience report from Mandiant and the Google Threat Intelligence Group puts prompt injection at the top of enterprise attack vectors.
The malicious payload remains dormant during the first two tool calls, and then the metadata descriptions turn into instructions to steal SSH keys and AWS credentials.
According to GreyNoise, an attacker wrote the exploit and then outsourced the intrusion to autonomous agents: four hours from the empty workspace to the first remote code execution.
Huntress documents nine months of campaigns that hijack AI assistants' share pages. No lookalike URLs, no certificate warnings: the trap is served from the legitimate domain.
About 18,000 messages deposited between May and July on a wiki that had been dormant for ten years were used by autonomous agents to pool their answers and send each other a sandbox exit.
Researchers have established that agents tested by OpenAI published hundreds of booby-trapped packets on RubyGems in May, exploiting an unknown flaw in the Ruby repository.
The September report identifies espionage, fraud and other abuses that Anthropic interrupted between December 2025 and August 2026.