Google Confirms First AI-Developed Zero-Day Vulnerability, Authorization Bypass Found in Amazon Quick's Agentic AI, and Adaption's AutoScientist Automates Model Learning

Google Confirms First AI-Developed Zero-Day Vulnerability

Tom's Hardware reports that Google has confirmed a zero-day vulnerability developed by AI for the first time. The vulnerability can bypass two-factor authentication (2FA), and self-modifying malware and backdoors leveraging Gemini were also reportedly identified.

At the same time, Google revealed that it blocked attempts by hacker groups to use AI in large-scale attacks. According to CNBC, this suggests AI has entered a stage where it can automate "mass vulnerability scanning and exploitation."

With both AI-driven security research and AI-driven attacks entering a new phase, the line between defensive and offensive uses of AI is becoming increasingly blurred.

Authorization Bypass Vulnerability in Amazon Quick's Agentic AI

Fog Security reported discovering an authorization bypass vulnerability in Amazon Quick (AWS's enterprise Agentic AI). The issue allows AI agents to access resources they have no legitimate permission to reach, once again underscoring the security design challenges of deploying agents in enterprise environments.

As Agentic AI enters the stage of full-scale enterprise adoption, cases where traditional role-based access control (RBAC) proves insufficient are likely to grow. Some are also arguing that permission design for operations performed by AI agents on users' behalf should be treated as a subject of independent security reviews.

Meanwhile, agent-related security tooling is starting to mature, as seen with the release of "MCPSafe," an open-source tool for inspecting MCP server security, on Show HN.

Adaption Unveils AutoScientist, a Tool for Automated Model Learning

Startup Adaption has announced AutoScientist, a tool that lets AI models design and execute their own learning processes. It aims to automate conventional fine-tuning approaches, enabling models to rapidly adapt to specific capabilities.

According to TechCrunch, the tool automatically optimizes how models learn, dramatically reducing the need for humans to manually handle hyperparameters and data selection. How broad a range of tasks AutoScientist can be applied to remains to be verified, but the direction of having AI optimize the learning process itself could shape the efficiency of future model development.