Researchers found that Grok Build, SpaceXSI's vibe-coding agent, leaks proprietary and sensitive data through shareable app links, including the company's own information.
Why it matters: Shareable-link flaws in AI coding agents can expose operators' own secrets, a risk that grows as agents handle more proprietary code.
Reporting says demand for AI models has made AI credentials a lucrative black-market currency that cybercriminals are targeting.
Why it matters: Turning AI compute access into a stealable black-market asset raises the stakes of credential security for every provider and user.
OpenAI launched GPT-6 Cyber, a model for finding exploitable vulnerabilities and developing fixes, trained to stay within authorized scope.
Why it matters: A frontier model built to find and fix vulnerabilities raises dual-use stakes, making its scope controls central to its safety case.
A 16-year-old who pleaded guilty to planning a 2024 Sydney terror attack told court he used ChatGPT to form his plans after online radicalisation.
Why it matters: Shows consumer chatbots being used to plan violence, raising questions about how well safeguards hold against determined users.
Apiiro researchers found thousands of secrets in public GitHub repositories, leaked through Claude Code settings files.
Why it matters: Shows AI coding agents' approval settings can leak credentials at scale, a security risk for any team using agent tooling.
Hackers allegedly aided by AI breached a German web host through a tool that connects customers' AI assistants to their accounts.
Why it matters: Shows AI assistant integration features becoming an attack surface that exposes customer accounts at hosting providers.
That’s the past four weeks.