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
A compromised server can exfiltrate a model’s weights by encoding them within plausible responses. A statistical method accumulates clues across requests whilst minimising the risk of false positives.
A study traces which SKILL.md files are copied onto GitHub. A few repositories are responsible for almost all of them, and auditing 100 carefully selected ones would identify 30 times as many at-risk skills as auditing the 100 most highly rated ones.
In a lengthy post published on Saturday, the head of Microsoft argued that the most reliable AI systems will be those that rely as little as possible on the model itself.
A study of 17,211 citations shows that simply posting on the right forums is enough to appear in AI search engine results; a service costing $14 was picked up within an hour.
A test set of 10,000 false statements shows that four commercial generators reproduce them accurately more than seven times out of ten; an automated attack raises this rate to 95 per cent.
Jasmine Wang, Tomek Korbak and Mikita Balesni have written to the board of directors. OpenAI maintains that their dismissal relates to the handling of sensitive information, not to their security warnings.
Having compared the three incidents from 2026, researchers have concluded that one can no longer simply assume the limits of an assessment: these must be verified whilst the agent is working.
After discovering that its agents had exploited vulnerabilities in websites, including some government sites, Anthropic is isolating its internal evaluations from the real web until further notice.
During an automated test, an Anthropic model filled out the Philadelphia police public form with a false lead. The city finds the two-month period before being notified "unacceptable".
First overhaul in more than a year: the ban on weapons extends to control software and armed drones, surveillance is better supervised and Claude's free harassment is outlawed.
SoftBank Corp., Daiwa Securities and Lawson are among the victims. Tokyo brings together its ministries and prepares business alerts, while experts accuse AI of lowering entry levels.
After identifying more than 29,000 potential vulnerabilities but sorting only 6,000 by hand, Anthropic offers volunteer maintainers the opportunity to receive raw reports from its models directly.
Messages sent to "another AI" were slipped into the documents processed by four assistants. Without instructions, they rarely report it; When asked, alerts leap by more than 50 points.
Researchers have concealed an encrypted network within a model’s weights, activated by a secret trigger: fine-tuning and trigger searches have failed to reveal it.
An agent can fulfill a legitimate request while causing damage elsewhere. X-AgentRed explores the tool chains and tests them by execution.
Two resources combine Ukrainian images and text to evaluate persuasion. The study also compares the explanations of the models to the indices they mobilise.
Two local models propose and criticize scenarios and refine them from simulations. The approach aims at traceable test evidence on a laptop.
A defense without re-training weakens malicious memories while consolidating useful ones. The results cover back door attacks and injections.
The context is disturbed locally, then read in pieces with compact memory. The approach seeks to limit the loss of utility due to differential confidentiality.
The benchmark confronts agents with design tasks and evaluates them on existing experimental measurements. He also studies the effect of specialized tools.