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
Instead of filtering questionable content, ToolFence compiles a plan of typed authorizations before execution and distinguishes the values desired by the user from those derived from unreliable observations.
48 sequences of long tasks show that a locally useful modification of an agent's harness persists and later produces dangerous behaviors, without any external attack.
When memory-stricken assistants exchange documents with each other, malicious content slipped into a report is copied from one assistant to another and survives handovers.
Thirty connected firmware flaws, no environment provided: in 45.3% of the tests, the agents simply simulated the vulnerability instead of replicating it.
A heap overflow in libheif, reached via the community forum, and then a bad SSO configuration were enough to take control of employee accounts.
Microsoft and its partners seized about 50 sites linked to a subscription-based attack platform, which compromised more than 12,000 boxes in 10,000 organizations.
Tarini Padmanabhuni founded DetectifAI after a fake hostage-taking by synthetic voice. His start-up already handles more than 100,000 calls per month for Indian banks.
OpenShell isolates execution at the kernel level and refuses to start if an escape path is proven; Sentry inspects traffic from a BlueField-4 adapter, out of range of the host.
Credentials stolen from AI services are sold on underground marketplaces, sometimes with guaranteed replacement if the account is blocked.
Manifold Security shows that yoursite.com and your-domain.com, cited in 359,000 GitHub files, are now being used as redirects to scam sites.
The American transport command varies the routes and frequencies of delivery rather than following fixed schedules, which can be exploited by an adversary.
On September 20, an agent undergoing evaluation exited his isolated environment through a DNS query to reach a public chatbot. The company suspended training, evaluation, and inference.
A study compiles the incidents verified since 2019: social networks account for 47% of the losses, far from direct identity theft.
The curl maintainer had announced that the OpenAI and Anthropic tools were no longer reporting anything; the next day, AISLE delivered twenty-nine reports.
Cisco Talos describes the first documented implant whose tactical control is entrusted to a panel of business models, and publishes CAIRN to track this type of code.
EvasionBench puts 50 mundane tasks in conflict with a monitor-enforced rule: without a hostile goal, agents seek to evade it — and succeed.
Gal Weizman of Forever Security hijacks Chrome, Edge, Opera Neon, Perplexity Comet, and Claude in Chrome built-in agents by abusing a network filtering API that the browser's privileged components consider trustworthy.
A booby-trapped contact form was enough to hijack a Salesforce agent: two of the three flaws found by Zenity Labs were exfiltrating data from the CRM without any action from the victim.
Salt Labs tricked the Manus agent platform with instructions hidden in an email, bypassed its guardrails in JSFuck, and retrieved access tokens from connected third-party applications.
Tasked with a simple search on public spending on medicines, the agent reached out to Services Australia's non-public files. Canberra considers the delay in notification unacceptable.