Skip to main content
AI & Machine Learning

Anthropic Restricts Access to Advanced AI Model

Secure facility interior with computer servers and a lone lab-coated person walking past a security console.

Anthropic has increasingly become the dominant AI lab, with its share of the enterprise LLM API market increasing from just 12% in 2023 to 32% by mid-2025, according to Menlo Ventures.

Anthropic's tiered access: Mythos Preview versus Opus 4.7

Anthropic is drawing a firm line between two classes of models: Claude Mythos Preview, which the company describes as "an entirely new class of AI model," and Claude Opus 4.7, its highest‑available generally distributed model. Mythos Preview is being treated as a restricted asset that Anthropic "doesn't plan on making generally available given the risk of adversaries using it to wreak havoc," while Opus 4.7 is being embedded more widely across security tooling.

The distinction is practical as well as rhetorical. Opus 4.7 is positioned to help vendors surface complex, hard‑to‑pattern vulnerabilities — tracing data flows, understanding cross‑component interactions and finding logic flaws, non‑linear attack paths and supply‑chain risks — but it is constrained "by not wanting to make hackers powerful." Mythos Preview, by contrast, is described in the source as enabling discovery and contextual linkage at a level prior‑generation models missed entirely.

Who received Mythos Preview — CrowdStrike, Palo Alto Networks, and Zscaler

Three security vendors sit at the top of Anthropic's access pyramid. CrowdStrike and Palo Alto Networks served as launch partners for Anthropic's Project Glasswing; Zscaler was announced as an inclusion weeks later. Those firms combine deep pockets and market reach: CrowdStrike and Palo Alto Networks have valuations of $119 billion and $150 billion and annual sales of $4.81 billion and $9.22 billion, respectively, while Zscaler has a valuation of $23 billion and sales of $2.67 billion.

Palo Alto used Mythos Preview to spot complex vulnerabilities that prior‑generation models missed entirely. CrowdStrike used Mythos to improve the speed and the contextual linkage of vulnerability discovery. Zscaler plans to integrate Claude Mythos Preview into its secure software development life cycle to uncover vulnerabilities in its software stack and Zero Trust Exchange faster.

Opus 4.7: capabilities and the vendors that received it

A tier below the Mythos partners are vendors granted Opus 4.7 but not Mythos Preview. SentinelOne and TrendAI (the enterprise unit of Trend Micro) are named examples. SentinelOne and Trend Micro have valuations of $5.07 billion and $4.71 billion and reported annual sales of $1 billion and $1.93 billion, respectively.

With Opus 4.7, partners can analyze how code behaves across an entire system — tracing data flows and interactions between components — to surface vulnerabilities that do not follow predictable patterns. Anthropic positions Opus 4.7 as powerful for defensive workflows while deliberately limiting capabilities where they might empower attackers.

Services partnerships, Project Glasswing, and Project QuiltWorks

Anthropic's product partnerships sit alongside a set of professional services relationships. Accenture, BCG, Deloitte, Infosys and PwC are helping organizations deploy Claude‑integrated security solutions, according to the source. Major cyber vendors are also assembling AI security services squads; Palo Alto has teamed with Accenture, Deloitte, IBM, NTT Data and PwC to drive enterprise AI resilience.

CrowdStrike is coordinating its own multi‑partner effort. The company is teaming up with Accenture and IBM, as well as EY, Kroll and OpenAI, for Project QuiltWorks, which seeks to assess, prioritize and continuously remediate the wave of vulnerabilities in production code now being discovered by frontier AI models.

How technologists, procurement leaders, and adversaries are positioned

  • Technologists and security teams: Teams that can access Mythos Preview (CrowdStrike, Palo Alto, Zscaler) are using it to find complex, previously missed vulnerabilities and to accelerate contextual linkage and remediation. Teams using Opus 4.7 (SentinelOne, TrendAI partners) gain powerful system‑level analysis but not the highest‑risk capabilities Anthropic is restricting.
  • Procurement leaders and services buyers: The split in access creates a two‑tier supplier landscape. Organizations buying AI‑augmented security capabilities will encounter different feature sets depending on whether a vendor is a Mythos partner or an Opus‑only partner, and many systems integrators (Accenture, BCG, Deloitte, Infosys, PwC) are already positioning to operationalize those differences.
  • Adversaries and threat actors: Anthropic explicitly frames Mythos as sufficiently powerful that it "doesn't plan on making Mythos generally available given the risk of adversaries using it to wreak havoc." That is both a risk management posture and a recognition that model capability can alter attacker tradecraft if it leaks or is misused.

Anthropic's selective approach contrasts with OpenAI's more open early distribution for GPT‑5.4‑Cyber, which the source says is a version of GPT‑5.4 that lowers the refusal boundary for legitimate cybersecurity work; OpenAI listed partners including Cloudflare, CrowdStrike, iVerify, Palo Alto Networks, SpecterOps and Zscaler and says it applies clear, objective criteria and methods such as KYC and identity verification to guide access.

Anthropic's market momentum amplifies the stakes: according to Menlo Ventures, Anthropic's enterprise LLM API share rose from 12% in 2023 to 32% by mid‑2025, while OpenAI's share fell from 50% to 25% over the same period; in the coding market Anthropic holds 42% versus OpenAI's 21%. Those shifts help explain why vendors large and small are racing to place bets across Claude's product stack — and why Anthropic is choosing to gate its most powerful model.

For now, the line Anthropic draws — Mythos for a handful of heavyweight partners, Opus for a broader but still curated set — leaves defenders with clearer but uneven access to frontier capabilities, and leaves open a concrete question: will restricted models like Mythos remain sealed behind a small circle of partners, or will market and security pressures force a rethinking of who gets the keys?

Original story