The Cyber Arms Race Has Two Sales Pitches

On April 7, Anthropic announced Claude Mythos Preview and Project Glasswing: an AI that autonomously finds zero-day vulnerabilities, restricted to roughly 50 vetted organizations. One week later, on April 14, OpenAI responded with GPT-5.4-Cyber and the Trusted Access for Cyber program, offering similar capabilities to thousands of verified defenders. The technology is converging. The access models are diverging.

What GPT-5.4-Cyber Does

GPT-5.4-Cyber is a fine-tuned variant of GPT-5.4 trained specifically for defensive cybersecurity.¹ The model handles binary reverse engineering, analyzing compiled executables to identify malware behavior and structural vulnerabilities without requiring source code. OpenAI deliberately lowered the model's refusal boundaries for security tasks, making it what they call "cyber-permissive" for legitimate defenders.

Access is gated through the Trusted Access for Cyber (TAC) program, which OpenAI introduced in February 2026. Individual defenders verify their identity at chatgpt.com/cyber. Enterprise teams apply through OpenAI account representatives. The stated goal: scale to "thousands of verified individual defenders and hundreds of teams responsible for defending critical software."²

No public pricing exists. GPT-5.4-Cyber is not in OpenAI's standard API pricing catalog. This is a restricted access program, not a product tier.

Two Philosophies

The contrast with Anthropic's approach is stark and deliberate.

Anthropic chose restriction. Mythos Preview goes to ~50 organizations: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, Nvidia.³ No general availability is planned. Anthropic committed a $100 million fund for securing critical software through Glasswing. The framing: these models are too dangerous for broad access and must be gated behind institutional trust.

OpenAI chose scale. TAC opens to thousands of verified defenders with a $10 million grant fund.⁴ The framing: broad access with identity verification is the only way to ensure defenders aren't outgunned by adversaries who face no such constraints. As OpenAI cyber researcher Fouad Matin put it: "This is a team sport. We need to make sure that every single team is empowered to secure their systems."⁵

The Schneier Take

Bruce Schneier isn't buying either pitch. On Anthropic's Glasswing: "The public has been given remarkably little with which to evaluate Anthropic's decision. We have been shown a highlight reel of spectacular successes."⁶ On the broader access question: "The claims that it's too dangerous to release publicly is pure marketing hype, just like OpenAI's nearly identical claim about GPT-2 was back in 2019."⁷

His deeper concern: "Anthropic is a private company and, in some ways, still a start-up. Yet it is making unilateral decisions about which pieces of our critical global infrastructure get defended first, and which must wait their turn." That critique applies equally to OpenAI's verification-gated model. Both companies are deciding who gets to defend what, and on what timeline. The difference is the radius of that decision.

What This Actually Is

Strip away the philosophical framing and the timeline tells the story. Anthropic announces an exclusive cybersecurity AI on April 7. OpenAI ships a competing product exactly one week later. The capability gap between the two companies is narrow enough that "who defends critical infrastructure" is now a go-to-market question, not a research question.

Both companies have economic incentives to position their model as the defender of critical infrastructure. Anthropic's restriction creates scarcity value. OpenAI's broad access creates market share. Neither company has released independent benchmarks comparing the two models' actual vulnerability discovery rates. We're comparing sales pitches, not capabilities.

Meanwhile, the UK's AI Safety Institute has warned that Claude Mythos Preview shows "unprecedented attack capability."⁸ Darktrace Federal CEO Marcus Fowler offered a useful reality check: "Most organizations are still constrained by the realities of remediation once an issue is discovered: patch development, testing, deployment, uptime requirements, and resource limitations."⁹

Finding vulnerabilities faster doesn't help if you can't fix them faster. That constraint applies regardless of which company's model finds them.