Skip to main content
AI Strategy & Workshops

April 7, 2026

When a Tech Company Chooses a Controlled Release

In Anthropic's own technical write-up, engineers describe pointing Mythos Preview at code before bed and waking up to working exploits.

Anthropic says Mythos Preview is a general-purpose frontier model, and the jump in cyber capability came from training a system that became exceptionally strong at code.

That is exactly why this matters: a better coding model also became a better vulnerability-discovery and exploit-development model.

Scroll

The Model

What Mythos Is

Anthropic calls Mythos Preview its most capable model yet. It announced the model publicly on April 7, 2026, and is limiting access to Project Glasswing participants.

Anthropic also makes the dual-use point explicit: Mythos Preview is a general-purpose model, and its cyber strength emerged as the system became exceptionally good at code. The same capability that helps with software development also helps with finding serious flaws in existing software.

The sharpest jump shows up in Anthropic's exploit benchmark. Opus 4.6 produced working Firefox exploits only twice in several hundred attempts. Mythos Preview produced 181 working exploits on the same benchmark and achieved register control 29 additional times.

0s

high-severity vulnerabilities already identified

27

years — oldest vulnerability discovered

0

working Firefox exploits in Anthropic's benchmark

Notable vulnerabilities discovered

OpenBSDSystem crash via network
27 years

Crash vulnerability hiding since 1999

FFmpegMedia processing exploit
16 years

Autonomously identified in a widely used media stack

FreeBSD 9.0Remote code execution
17 years

Anthropic reports full remote-code-execution exploitability

Anthropic describes Mythos Preview as a general-purpose frontier model whose cyber strength emerged from becoming exceptionally good at code.

Anthropic, Project Glasswing announcement

The Release Decision

Why Access Is Controlled

Anthropic built what it calls its most capable model yet and chose a tightly controlled release.

Anthropic treated this as a controlled release and routed access through Project Glasswing, a coalition built to find and fix vulnerabilities in critical infrastructure.

Anthropic's stated rationale

Anthropic says Mythos Preview crossed a capability threshold: researchers using it have already found thousands of high-severity vulnerabilities, including some in every major operating system and every major web browser, which is why the company is keeping access inside Project Glasswing.

Anthropic's technical write-up says Mythos Preview can now find and help weaponize serious software flaws at a scale prior frontier models never reached. That is enough on its own to change the release calculus.

Anthropic says fewer than 1% of the vulnerabilities Mythos Preview surfaced had been fully patched by maintainers when it published the technical write-up. That creates a clear math problem for defenders: patching and response speed now have to rise with discovery speed.

Anthropic's disclosed findings

Patching status

<1%

fully patched

UnpatchedPatched

The Landscape

What This Signals About the Industry

A few things to keep in perspective here.

From early February to early April, Anthropic, Google, OpenAI, and xAI all announced major frontier-model releases or upgrades. Anthropic's outlier move was keeping its newest coding model gated.

Frontier models — April 2026

AnthropicOpus 4.6

Agentic coding upgrade

Feb 2026released
GoogleGemini 3.1 Pro

Frontier reasoning and coding update

Feb 2026released
OpenAIGPT-5.4

Frontier release focused on coding

Mar 2026released
xAIGrok 4.20 beta

Beta release in xAI API

Mar 2026beta
AnthropicMythos Preview

No general release; gated via Glasswing

Apr 2026withheld

Five major coding-model announcements in about two months. Anthropic is the one that stayed gated.

The more important signal is Anthropic's claim that the cyber jump came from a general-purpose coding model. That raises a broader industry question about where powerful coding models are heading.

The surrounding response is also notable. Project Glasswing launched with 12 named partners and Anthropic says it is extending access to 40+ additional organizations that build or maintain critical infrastructure. The partner list spans cloud providers, chip makers, banks, and security firms, which suggests they view this as an operational issue with immediate consequences.

We've reached the point where a frontier lab is making release decisions around exploit capability at all.

The Implications

What This Means for Your Organization

The practical question for a business leader: what do we do about it?

Assume proliferation

0%

of organizations reported breaches of AI models or applications

Know your data

0%

of organizations lack or are unsure they have the right data management practices for AI

Shadow AI risk

+$0K

added to average breach costs when shadow-AI exposure is high

Anthropic's write-up is unusually explicit on the direction of travel: the techniques involved are well-understood primitives, and the company says the current trend is likely to continue. For a business leader, that makes governance, visibility, and response speed more urgent than trying to guess which lab gets there next.

Your Move

Three Questions Worth Answering Now

These are operational questions with concrete answers. Most teams feel the gap between knowing what to do and having the bandwidth to do it.

What Comes Next

Where This Goes

The defender's dilemma is asymmetric. Defenders patch every vulnerability. Attackers look for one opening. AI pushes more speed and scale toward offense, so defensive deployment needs a meaningful head start.

Project Glasswing

$104M

committed to collective AI defense

12

launch partners

40+

additional organizations

AnthropicAWSAppleBroadcomCiscoCrowdStrikeGoogleJPMorganChaseLinux FoundationMicrosoftNVIDIAPalo Alto Networks+40 additional orgs

Anthropic says it will share what it learns from Glasswing so the broader software ecosystem can benefit.

Anthropic says Mythos Preview's cyber lift arrived as a downstream consequence of the same improvements in coding, reasoning, and autonomy that make the model useful for legitimate software work.

Anthropic's write-up argues the underlying building blocks are now well-understood and that the current trend is likely to continue. These tools are already here. The practical question is who gets to use them first and under what controls.

Anthropic's Mythos Preview disclosures point to a new phase: frontier coding models can now surface and sometimes weaponize serious software flaws at a pace that changes how organizations need to govern AI, patch systems, and monitor access.

Sources

Verified References

This page was reviewed against the public materials below on April 8, 2026. The copy above now tracks these sources closely and uses their level of precision.

Anthropic — Project Glasswing

Primary source for why Anthropic is gating Mythos Preview, the 12 named launch partners, 40+ additional organizations, and the $100M + $4M Glasswing commitment.

xAI — Release Notes

xAI's March 10, 2026 Grok 4.20 beta release note used for the release-timeline comparison.