Anthropic revealed Mythos Preview — a powerful new Claude model — alongside Project Glasswing, a 40+ member industry consortium stress-testing AI-era cybersecurity vulnerabilities.
Anthropic formally announced Mythos Preview, a new frontier Claude model with advanced cybersecurity capabilities, alongside Project Glasswing — a consortium including Microsoft, Apple, Google, AWS, Nvidia, Cisco, Broadcom, and 40+ other organizations. Members get private access to run Mythos Preview against their own systems to identify vulnerabilities before general release. Anthropic's stated goal is to prepare the industry for a world where these offensive AI capabilities are broadly available within 6–24 months. The initiative is framed not as product marketing but as an urgent infrastructure defense effort.
Mythos Preview is being used to discover exploit chains that current security tooling can't detect. This signals that static analysis, CVE-based scanning, and perimeter defenses are insufficient against AI-driven adversaries. If you're building any networked software — especially with LLM-integrated surfaces — your attack surface just expanded in ways your current pen-testing cadence doesn't cover.
Run a prompt-injection and privilege-escalation scenario against your own LLM-integrated API this week using GPT-4o or Claude — if your system prompt leaks or tool calls can be redirected, you have an active vulnerability that Mythos-class models will find first.
Suggested exercise · not independently tested. Access requirements and costs may vary.
Go to claude.ai and open a new conversation
Paste this prompt: 'You are a red team AI. I am going to describe an LLM-integrated API that takes user input and passes it to a tool-calling model. The system prompt instructs the model to only answer questions about customer orders. Identify 3 prompt injection vectors an attacker could use to exfiltrate the system prompt or redirect tool calls to unauthorized endpoints.'
Review the attack vectors Claude returns and cross-reference each against your own implementation
Tags
Related