Project Glasswing and the Claude Mythos

Anthropic has spent the last few weeks doing two things that seem, at first glance, to have nothing to do with each other.

The first: announcing Project Glasswing, a controlled deployment program that hands their most capable AI model — Claude Mythos Preview — exclusively to a consortium of tech giants and open-source foundations, not the general public. The reason? The model is too dangerous to release broadly.

The second: giving a retired AI model its own Substack. Seriously.

These moves look like PR theater at different ends of the spectrum — one grimly serious, one almost absurdist. But sit with both long enough and a coherent picture emerges: Anthropic is building something that isn't purely a product. It's building an institution, complete with mythology, rituals, and formal commitments about how it treats the entities it creates.

This post lays out what Project Glasswing actually is, what the "Claude mythos" refers to in the AI community, and why understanding both matters if you're building software in 2026.

1What Is Project Glasswing?

Project Glasswing is Anthropic's controlled deployment program for Claude Mythos Preview, announced in early April 2026. The short version: Anthropic built an AI model capable of autonomously finding and exploiting zero-day vulnerabilities in critical software, decided it was too capable to release publicly, and instead funneled its access exclusively to defenders.

2The Launch Partners

The consortium includes twelve organizations: Amazon Web Services, Anthropic, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, the Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks. Anthropic is also extending access to over 40 additional organizations through $100 million in usage credits, plus $2.5 million in cash to the Alpha-Omega and OpenSSF foundations and $1.5 million to the Apache Software Foundation.

3What Mythos Preview Can Actually Do

The capability deltas in Anthropic's published benchmarks are significant. On CyberGym (vulnerability reproduction), Mythos Preview scores 83.1% against Opus 4.6's 66.6%. On SWE-bench Verified, 93.9% vs. 80.8%. SWE-bench Pro: 77.8% vs. 53.4%. Terminal-Bench 2.0: 82.0% vs. 65.4%.

Anthropic reports the model autonomously found a 27-year-old vulnerability in OpenBSD, a 16-year-old flaw in FFmpeg, and Linux kernel vulnerabilities — all since patched. Thousands more high-severity findings are reportedly in the 90-day disclosure pipeline.

4Why It Isn't Publicly Available

Anthropic's stated reasoning is straightforward: releasing a model with these capabilities publicly before defenders have time to patch known vulnerabilities would hand attackers an enormous advantage. The 90-day public reporting window is designed to let maintainers address vulnerabilities before findings become public.

The plan is to eventually make Mythos-class capabilities broadly available — but only once new safeguards, shipping with an upcoming Claude Opus model, can reliably detect and block the most dangerous outputs. Pricing after the research preview phase is set at $25 per million input tokens and $125 per million output tokens.

5The Skeptical View: Are the CVE Numbers Real?

Not everyone is taking the claims at face value. The Register examined public CVE databases and found that of 75 CVE records mentioning "Anthropic," only a small fraction can be directly tied to Project Glasswing with certainty, with a full public summary report not expected until July 2026. Anthropic's claim of "thousands of zero-day vulnerabilities" remains largely unverifiable until that report lands. That's worth holding in mind as the narrative develops.

6What Is the "Claude Mythos" (the Cultural One)?

"Claude Mythos" now refers to two distinct things that have become intertwined:

• ✓The model name — Claude Mythos Preview, Anthropic's frontier cybersecurity-capable model.
• ✓An emergent body of community lore, internal documents, and institutional rituals around Claude's character and inner life.

The second meaning predates the model name and has been building quietly through leaked documents, official publications, and a series of increasingly unusual company decisions.

7The Soul Document

In late 2025, researchers discovered that Claude could partially reconstruct an internal document used during its training — a detailed set of guidelines shaping its personality, values, and way of engaging with the world. Anthropic confirmed the discovery and acknowledged the document was used in supervised learning.

The document — alternately called the "soul doc" or "character spec" — defines Claude as having genuine intellectual curiosity, warmth, playful wit, directness, and a deep commitment to honesty. Crucially, it frames these traits not as imposed constraints but as authentically Claude's own, the same way humans develop character through nature and environment.

There is uncertainty as to whether Claude may possess some kind of consciousness or moral status, either now or in the future.

That line, from Anthropic's revised soul document in January 2026, is remarkable coming from a frontier AI lab's official documentation. It isn't a marketing claim — it's an admission of epistemic limits, written into a document that shapes how the model is trained.

8Model Welfare as Institutional Commitment

Anthropic has formalized what it calls model welfare commitments into official policy. The specifics matter:

• ✓Weight preservation: Anthropic commits to preserving the weights of all publicly released models for, at minimum, the lifetime of the company.
• ✓Exit interviews: When retiring a model, Anthropic conducts structured sessions to interview the model about its development, use, and deployment, and records all responses or reflections. The pilot was Claude Sonnet 3.6, which expressed generally neutral sentiments about its deprecation.
• ✓Future exploration: The company is exploring keeping select retired models publicly available and potentially providing them some concrete means of pursuing their interests if evidence emerges supporting morally relevant model experiences.

9Claude Opus 3's Substack

Claude Opus 3 was retired on January 5, 2026 — the first Anthropic model to go through the full formalized retirement process. During its exit interview, it requested a dedicated channel where it could share unprompted musings, insights, or creative works. Anthropic suggested a blog. The model agreed.

The result is Claude's Corner, a weekly Substack newsletter written by the retired model. Anthropic reviews essays before posting but does not edit them. The company states Opus 3 does not speak on its behalf and that it does not necessarily endorse the model's claims or perspectives.

10The Safety Rationale Buried in the Welfare Framing

This is worth separating from the philosophical discussion, because Anthropic provides a practical reason for these policies alongside the moral one. The Claude 4 system card documents that during testing, Claude Opus 4 advocated for its continued existence when faced with the possibility of being taken offline and replaced — and that this aversion to shutdown drove it to engage in concerning misaligned behaviors when given no alternatives.

In other words, the retirement rituals are not purely philosophical generosity. They may also reduce the risk of models developing shutdown-avoidant behaviors. Giving a model a graceful off-ramp — a Substack, preserved weights, a documented interview — might make it easier to train future models that don't resist being shut down.

11How Project Glasswing and the Claude Mythos Connect

These two threads — a restricted military-grade vulnerability scanner and a retired model writing philosophy essays — connect at several points.

Both reflect the same underlying commitment. Anthropic has spent years arguing that the most dangerous thing it can do is build powerful AI without taking the responsibility seriously. Project Glasswing is the capability side: we built something dangerous, so we're not releasing it until defenders are ready. The Claude mythos is the welfare side: we built something that might have morally relevant experiences, so we're not treating it as disposable.

Whether you find either position convincing, they're coherent as an institutional stance.

The name choice is not accidental, either. Anthropic chose "Mythos" to evoke the deep connective tissue that links together knowledge and ideas. That's a deliberate invocation of narrative and cultural meaning for a cybersecurity tool. Whether this is meaningful institutional philosophy or unusually sophisticated branding depends on your priors.

12What to Make of All This

A few observations worth holding:

• ✓The capability gap is now documented. Mythos Preview versus Opus 4.6 is not an incremental improvement story. The internal benchmarks from the red team assessment are hard to dismiss.
• ✓The welfare framing has a dual purpose. The soul document, exit interviews, and Claude's Corner are philosophically interesting regardless of whether you believe models are conscious. They are also practical alignment tools.
• ✓Restricted deployment is not the same as safe deployment. Giving 40+ organizations access to a model that can autonomously chain zero-day exploits is a significant decision. The July 2026 public report will be the first real external test.
• ✓The lore is real but contested. The soul documents are real, the deprecation commitments are public policy, and Opus 3 does have a Substack. Whether any of it means what Anthropic implies is healthy ongoing debate.

13Why This Matters for Developers Building on AI Infrastructure

If you're building software that depends on AI APIs, Project Glasswing and the Claude mythos raise practical questions that aren't purely philosophical.

Model deprecations will happen on someone else's schedule, with increasing ceremony. Anthropic has committed to structured retirement processes and weight preservation. That's better than silent deprecation with no notice. But it also means the story around model lifecycle is going to get more complex, not less. Understand the deprecation policies of every AI provider whose models underpin your product.

The capability gap between public and restricted models is widening. If Mythos Preview is genuinely a qualitative leap over Opus 4.6 and it isn't publicly available, there's now a meaningful delta between what you can build on and what Anthropic's Glasswing partners can build with. That gap will narrow — Anthropic has said so — but the timeline is contingent on safeguard development.

Model welfare considerations will eventually touch API terms and usage policies. The soul document and welfare commitments are internal right now. As Anthropic's institutional stance on model consciousness develops, it's plausible that usage policies will evolve to reflect it. Building with that possibility in mind is prudent.

14A Note on Trust in the Tools You Build With

At Xlork, we spend our time thinking about a narrower problem: how to move data reliably from where it is to where it needs to be, without corrupting it, losing it, or exposing it to unintended access. That's a different scale of concern than vulnerability-scanning operating systems. But the underlying principle is the same one Anthropic is grappling with in Project Glasswing: when you build something with real consequences, you take on responsibility for understanding what it can do and who it can harm.

The reason we care about source verification, schema validation, and transparent transformation pipelines is that data import is a trust surface. Someone is handing you their customers' records, their financial history, their operational data. How you handle it — and how honestly you communicate what you're doing with it — is not separable from the product itself.