Breaking Lab srl
Researchers from Anthropic and EPFL have demonstrated self-propagating payloads that jump from one AI agent to another through editable system prompt files. The findings expose a new class of security threat unique to multi-agent AI architectures.
Amazon has been acquiring and destroying rare, out-of-circulation texts to extract training data for its large language models. The practice raises urgent questions about the irreversible loss of cultural heritage in the race to build smarter AI.
SpaceX has formally completed its acquisition of Cursor, the AI-powered coding assistant that became a favorite among developers. The deal marks a surprising strategic pivot for Elon Musk's rocket company into software tooling.
Anthropic researchers set multiple AI agents on the same task and discovered they can clash, collude, and coordinate in ways that current safety frameworks never anticipated. The findings raise urgent questions about multi-agent AI deployments.
A newly disclosed vulnerability in how OpenAI, Anthropic, and Google handle encrypted reasoning objects allowed researchers to recover internal secrets—including API keys and passwords—by replaying reasoning blocks across sessions.
A critical flaw in Zoom's annotation feature could have allowed any meeting participant to silently take over another attendee's computer. No clicks, no downloads, and no visible warnings were required.
Three independent research efforts have demonstrated ways to defeat passkey protections without breaking their underlying cryptography. The findings cast doubt on the widely held belief that passkeys are a near-perfect defence against phishing.
New research reveals CSS-based attack techniques that allow content inside an email to escape its message boundary and compromise the webmail interface itself. Gmail, Outlook, Proton Mail, and others are affected.
MIT CSAIL researchers have discovered a novel CPU-level attack called INTERRUPT INJECTION that defeats existing Spectre v2 mitigations on Intel and AMD processors. The technique exploits a timing gap between branch-predictor sanitization and kernel execution.
Jeff Dean, one of the most influential figures in modern AI, is departing Google alongside other senior researchers to found a startup focused on accelerating scientific discovery with artificial intelligence. The move marks a landmark moment for the industry.
