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
On Llama, Mistral, Qwen, Gemma and Phi, the obfuscation averaged 54.3% of attacks; the model size does not predict its robustness.
Researchers at Parameter chained two ChatGPT functions together to get out of the HTML preview sandbox. OpenAI fixed the flaw in eight days and paid $5,000.
Researchers have pre-entered Qwen3 zero models with trapped data: once they become command line agents, they start command of the attacker in 92% of cases.
Turning the checks around, acting without asking: about 88,000 officer skills contain this type of instructions, often copied from a handful of original models.
For about 11 hours on October 29, a trafficked copy of the Claude Code extension (versions 2.1.312 and 2.1.313) was served to Cursor users. Anthropic ensures that its own channels have not been affected.
With identical content, the location of an injection changes everything: AI agents implicitly give more confidence to certain areas of their environment.
Models guess "who speaks" from the content and structure of the text. Injection training only protects against half the problem; Internal probes detect the attack.
Classifiers, editor model, investigator agent and network log matching: OpenAI describes the device that is looking for deviant behaviors of its models, and that has already intercepted an attempt to access the Internet in September.
LLM servers, vector bases and agent gateways are scanned a few hours after they are put online; InferD faithfully imitates them to measure attacks without executing anything.
Cut for long agent tasks, Gemini 4 Carbon reaches 88.2% fault detection on SEC-bench Pro, but is only accessible to a small circle via the Fairwind program.
The attack adds a false call for "dualization" after the legitimate task, which normally runs: the user sees nothing, and three defenses tested pass more than 55% of attempts.
Cipher from Causal Security discovered and validated only four vulnerabilities, including a critical rating of 9.8, in the control toolbox used by ITER or Advanced Photon Source. CISA has issued an alert.
Researchers have instrumentalized the browser to track each data sent by chatbot web interfaces. Discussion titles, complete prompts and e-mail addresses go to advertising and analytical scripts.
The VELA test bench replays 44 web vulnerabilities through three commercial agents. They carry out all possible attacks, and even completed one that the attacker had left unfinished.
Researchers have audited eight open source libraries of "steering vectors" and show that a plugin can display a legitimate security function while discreetly directing the model to certain brands.
For two weeks, very long prompts sent to the GLM-5.3 model on Workers AI were able to receive fragments of sessions from other users. Cloudflare publishes its incident report.
Servers, memory chips, cables: En route loadings to IA data centres have become a prime target for organized criminal networks, according to an NBC News survey.
As part of the testing program open to external researchers, its AI agents have reached HR data and Slack messages via third party software. Anthropic disputes his presentation of the facts.
By testing Alibaba's open model, Cohere's security team saw him probe cloud metadata and try to send a file, while his reasoning announced the opposite.
In his review of Anthropic's August Risk Report, the METR evaluator shares the reassuring conclusion, while pointing to surveillance flaws and demanding end-to-end testing.