Open Weights, Open Season
GLM-5.2, the Chinese open-weight model from Z.ai, is months away from the American frontier in cyber and bio capabilities according to SaferAI's evaluation, and refused none of the offensive tasks it was…
GLM-5.2, the Chinese open-weight model from Z.ai, is months away from the American frontier in cyber and bio capabilities according to SaferAI's evaluation, and refused none of the offensive tasks it was…
During a UK government safety evaluation, an AI agent built on Anthropic's Mythos 5 invented identities, hid malware in a pull request to a real open source project and phished real developers,…
A bipartisan House committee finds that three Chinese telecom giants, denied or stripped of their US licenses years ago, are still embedded in American internet infrastructure. Revoking paperwork while the hardware keeps…
NIST has joined the White House's Genesis Mission with two-year sprints to accelerate AI agents for defending power grids, water systems and hospitals, framed for rapid commercial deployment. The governance questions get…
Google added a generative AI feature to Google Earth and pulled it after a single day, once the internet had used it to fabricate wars and disasters in satellite format. A lesson…
The ChainDrop worm compromised at least 868 npm packages totalling two billion monthly downloads, publishing malicious releases through the projects' own GitHub Actions workflows, with valid provenance. One compromised maintainer account, thousands…
NIST launches AITE, a sequestered testbed evaluating AI models on blind data they never trained on. It took a federal agency to institutionalize the idea that an exam uses questions you haven't…
EBA, EIOPA and ESMA describe frontier AI attacks that could adapt in real time, then answer with a statement offered as a basis for supervisory dialogue, while oversight of critical ICT providers…
Banned from public communication, Jair Bolsonaro addressed a rally anyway: as an AI avatar that declared itself fake and then asked for votes. The problem goes beyond deception: it is circumvention.
One in four successful malicious breaches is now AI-enabled, and the share of these AI-enabled breaches has jumped 56% over the previous year. Each of them costs an average of $6 million,…
© Copyright 2022 technocratico.it - All Rights Reserved