Anthropic's Mythos model — described by testers as 'a generational leap in AI capability' — remains under restricted access through Project Glasswing, with only approximately 50 vetted partners including Amazon, Microsoft, Apple, Google, and Nvidia able to evaluate it. Internal benchmarks suggest Mythos represents the largest capability jump since GPT-4, particularly in cybersecurity, advanced reasoning, and autonomous code generation. The NSA has begun evaluating Mythos for identifying and patching network vulnerabilities despite the Pentagon's ongoing supply-chain risk designation against Anthropic.
The paradox has not gone unnoticed. Pentagon CTO Emil Michael acknowledged on May 1 that while the blacklist 'stands,' Mythos access is 'a separate national security moment.' The White House is reportedly drafting executive action on Mythos access. The situation has created a de facto two-track policy: Anthropic is too dangerous to do business with for general-purpose AI, but its most advanced model is too useful not to evaluate for the highest-stakes national security applications.
Cerebras raised its IPO price range from $115-$125 to $150-$160 per share and increased the share count from 28 million to 30 million after institutional demand flooded the order book. At the top of the new range, the offering raises approximately $4.8 billion. The amended S-1 filed May 4 reflected the updated terms, and the listing is scheduled for May 14 on Nasdaq under ticker CBRS. Reuters sources described the order book as more than 20 times oversubscribed.
The strong demand validates investor appetite for AI infrastructure plays beyond NVIDIA. Cerebras's $510 million 2025 revenue, $24.6 billion backlog, and $10 billion OpenAI deal provide a revenue visibility uncommon in the chip industry. The company's wafer-scale architecture — a single chip 57x larger than NVIDIA's H100 — has attracted particular interest from hyperscalers seeking alternatives to GPU clusters for inference workloads. Amazon Web Services has entered as a distribution channel through AWS Bedrock.
OpenAI released GPT-5.4-Cyber, a specialized variant of its flagship model with reduced restrictions on cybersecurity tasks, on April 15. The model is available only to vetted security vendors, organizations, and researchers approved for the highest access tier. It has fewer constraints on sensitive cybersecurity activities including vulnerability research, exploit analysis, and penetration testing — tasks that standard GPT-5.4 refuses to perform due to safety policies.
The release marks OpenAI's entry into the 'offensive-defense' AI category, following Anthropic's Mythos and Project Glasswing. GPT-5.4-Cyber shares the same base architecture as GPT-5.4 but removes specific refusal patterns for cybersecurity queries while adding domain-specific knowledge for vulnerability research. OpenAI has implemented monitoring and usage tracking to detect abuse. Early testers report the model can identify zero-day patterns in source code and suggest exploitation paths that human researchers would take days to discover.