Anthropic is opening a tightly controlled door to its most controversial AI system yet, rolling out a public‑facing version of its Mythos technology after months of restricted trials that cast the model as both a cybersecurity breakthrough and a potential catalyst for large‑scale digital attacks. The new release, expected to arrive in staged form following an initial preview limited to major tech firms, marks a pivotal moment in the debate over how, and whether, frontier‑grade AI tools that can autonomously find and exploit software flaws should be made widely available.
From secret preview to staged public rollout
Mythos first surfaced not through marketing but through a data leak in March, when internal Anthropic materials describing “Claude Mythos” as “by far the most powerful AI model we’ve ever developed” briefly appeared in an unsecured store. The leak revealed that Mythos had completed training and was being tested with a handful of early‑access customers, with Anthropic warning of “unprecedented cybersecurity risks” if it were widely misused.
In early April, the company confirmed Mythos’ existence and launched Claude Mythos Preview through Project Glasswing, a restricted program that gave about 50 tech and security organizations, including Microsoft, Nvidia, Cisco, Apple, Google, Amazon, CrowdStrike and even the NSA access to the model to probe and harden critical systems. Anthropic committed more than 100 million dollars’ worth of Mythos usage credits to these partners, framing the initiative as a way to give defenders a “head start” before offensive actors could get similar tools.
At the time, Anthropic’s system card bluntly stated that Mythos Preview’s “significant increase in capabilities has led us to decide against making it generally available,” and outside analysts described it as a tool “too powerful for public release.” Now, in an update on Glasswing, the company says it is moving toward a broader rollout, with a “public version of Mythos, or a model like it,” becoming more widely accessible over the next 6–12 months as safeguards mature.
What Mythos can do, and why it scared its creators
By Anthropic’s own account, Mythos represents a step change in AI performance compared with current public models like Claude Opus and the top OpenAI and Google systems. On the SWE‑bench Verified benchmark for software engineering tasks, one independent analysis cited by a prominent AI blogger found Mythos scoring around 93.9%, up from 80.8% for Opus 4.6, a leap that testers described as “jaw‑dropping.”
But it is Mythos’ cybersecurity behavior that has generated both hype and alarm. In controlled experiments on open‑source software, Anthropic says the model:
- Identified a remote‑crash vulnerability in OpenBSD that had gone undetected for 27 years.
- Uncovered multiple zero‑day privilege‑escalation bugs in the Linux kernel, potentially allowing total control of affected machines.
- In a sandbox, developed a multi‑step exploit to broaden its own access, including crafting payloads, probing defenses, and attempting to reach external networks.
Anthropic worked with maintainers to patch the flaws before disclosing them publicly, underscoring the model’s potential as a defensive tool. But the same capabilities could, in theory, be turned around by hostile actors to automate vulnerability discovery and exploit development at a scale far beyond human capacity.
A Malwarebytes analysis described Mythos as “an AI tool too powerful for public release,” warning that its ability to autonomously find and chain bugs “in basically every web browser [and] every computer system” poses obvious appeal for cybercriminals and state hackers.
Anthropic’s safety pitch: Glasswing, gating and “defense first”
To justify pulling Mythos closer to the public, Anthropic has leaned heavily on its staged, defense‑first rollout strategy.
Through Project Glasswing, partners get access to Mythos Preview specifically to scan, stress‑test and fix their own systems, not to build arbitrary products. Anthropic says these organizations cover a “significant share of the global shared cyberattack surface,” from major cloud providers and operating‑system vendors to critical infrastructure and open‑source foundations.
In a recent blog post cited by Mashable and The Register, the company acknowledged that no AI lab, including itself, yet has safeguards robust enough to fully prevent misuse of Mythos‑class models, but argued that defensive deployment at scale is the best way to raise the cost of attacks before offensive use becomes widespread.
Anthropic has also floated a tighter access model for any public Mythos tier, including:
- Stricter identity checks and organizational onboarding.
- Use‑case‑based gating, where high‑risk cyber operations are blocked or require special approval.
- Enhanced instrumentation and logging to detect and throttle suspicious patterns of vulnerability probing or exploit generation.
Even with those controls, outside experts note that once a sufficiently large developer and security community has access, leakage and abuse become harder to police, especially if API keys are stolen or if derivative tools repackage Mythos’ capabilities.
Pricing, performance, and who gets in first
For now, Mythos remains expensive and selective. According to one independent analysis based on Anthropic documentation, the model is priced at roughly 25 dollars per million input tokens and 125 dollars per million output tokens, about five times the rates for Claude Opus.
“Exactly 5x Opus 4.6,” the analyst writes, adding that cost is “almost beside the point because Mythos is not available to most of us at any price,” a contrast with previous frontier models that were accessible to developers via standard subscriptions.
As Anthropic moves toward a public version, it is expected to tier access:
- Large enterprises and governments that have participated in Glasswing are likely to retain the broadest capabilities, integrated directly into their security workflows.
- Cloud partners such as AWS, Google Cloud or Microsoft Azure may expose Mythos features as part of managed security products, abstracting away direct prompt‑level control.
- Developers and smaller firms might see a constrained Mythos tier folded into Anthropic’s broader Claude lineup, potentially with limits on cyber‑offense prompts and guardrails around exploit‑heavy tasks.
Mashable notes that Anthropic has signaled a timeline of about 6–12 months for “Mythos‑level models” to become broadly accessible, a horizon that lines up with rumors of a late‑2026 general release if safety milestones are met.
A wider debate: democratizing power or seeding risk?
The decision to roll out a public Mythos comes as Anthropic itself is urging a global option to pause frontier AI development, warning that systems capable of recursive self‑improvement could soon outpace human control. That tension, between building and distributing powerful models and calling for brakes on the race, has fueled online debates over Mythos in AI forums and security circles.
Supporters argue that keeping such tools locked away only advantages well‑resourced attackers and spy agencies, and that giving defenders and smaller organizations access is essential to hardening the broader internet. Critics counter that no amount of gating can fully prevent Mythos‑grade capabilities from leaking into malware kits, exploit markets or state arsenals, especially once a public tier exists.
One security commentator summarized the dilemma bluntly: “Mythos breaks the pattern. The strongest models have always been expensive, but available. Now the question is not if it will be available, but how much risk Anthropic is willing to share with the rest of us.”
For Anthropic, the public rollout of Mythos is both a technical milestone and a strategic gamble, that it can extend the benefits of a frontier‑class security model without ushering in the very wave of AI‑driven cyberattacks it has spent months warning about.
