The AI Arms Race: OpenAI Counter-Programs Anthropic’s "Mythos" with New Cyber Pilot

As the industry shifts toward "AI-First" defense, the two titans of Silicon Valley are taking different paths to the same goal: automating the discovery and patching of global software vulnerabilities.

SAN FRANCISCO, CA — The battle for the future of cybersecurity has entered a restricted phase. Following the bombshell reveal of Claude Mythos — an Anthropic model reportedly capable of finding "thousands of zero-day vulnerabilities" autonomously — OpenAI has finalized its counter-move: the "Trusted Access for Cyber" pilot program.

While Anthropic has dominated headlines this week with the terrifying efficacy of Mythos, OpenAI is quietly readying a specialized rollout of its GPT-5.3-Codex architecture. Both companies are now treating their most advanced models less like commercial chatbots and more like digital weaponry, locking them behind "gated" programs for a select group of defenders.

Anthropic's Mythos and Project Glasswing

Anthropic’s newest frontier model, Claude Mythos, has sent shockwaves through the security community. In early testing, the model identified high-severity zero-day vulnerabilities across every major operating system and web browser.

In response, Anthropic launched Project Glasswing, a massive defensive alliance. Unlike a standard product launch, Glasswing provides exclusive access to Mythos for a "vetted circle" of partners, including Google, Microsoft, AWS, Cisco, and JPMorgan Chase. The initiative includes $100 million in usage credits to help these organizations secure the world's most critical open-source and enterprise software.

OpenAI’s "Trusted Access" Blueprint

OpenAI is executing a parallel strategy through its "Trusted Access for Cyber" program, first initiated in February and expanded this week. This program provides controlled access to GPT-5.3-Codex, a model OpenAI recently designated as "High Capability" in the cybersecurity domain.

To ensure the model is used for defensive acceleration rather than offensive exploitation, OpenAI is backing the program with $10 million in API credits for participating security researchers. The goal is to harden "baseline safeguards" across the internet before these capabilities become available to threat actors.

The CyberSignal Analysis

Signal 01 — The Verticalization of AI

The launch of these programs confirms that general-purpose AI is bifurcating. OpenAI and Anthropic are now verticalizing their security offerings, acknowledging that standard models lack the specialized logic required for high-stakes vulnerability discovery.

Signal 02 — The Security Divide

A "Security Divide" is emerging. Organizations inside these gated programs will become exponentially more resilient, while SMBs and projects not included in the vetting may become the primary targets for threat actors who eventually replicate these capabilities.

Sources