Particle.news
Download on the App Store

OpenAI Launches GPT-5.4-Cyber as Anthropic’s Mythos Remains Under Tight Preview

Regulators warn that AI models that hunt software flaws compress the time from discovery to exploitation.

Overview

  • OpenAI, which unveiled GPT-5.4-Cyber on Tuesday, is expanding its Trusted Access for Cyber program to verified defenders and giving them a fine-tuned model for tasks like vulnerability research and binary reverse engineering after identity checks.
  • Anthropic continues to limit Claude Mythos to a select consortium under Project Glasswing as it claims the model has uncovered thousands of previously unknown flaws across major operating systems and web browsers.
  • Treasury Secretary Scott Bessent and Federal Reserve Chair Jerome Powell briefed bank chiefs on the risks tied to Mythos, signaling concern that financial networks could face faster, AI-driven exploit chains.
  • Security experts say these systems are dual-use and lower the barrier to finding exploitable bugs, with Rich Mogull urging organizations to adopt AI-assisted testing and speed up patching to keep pace.
  • Independent teams reproduced some Mythos-highlighted vulnerabilities on smaller, cheaper models, suggesting the capability is spreading and pushing vendors to control access, verify users, and harden safeguards.