LumiChats first covered Project Glasswing in April 2026, when Anthropic said it had built a cybersecurity model too capable to release publicly and instead handed early access to Apple, Google, Microsoft and a handful of other tech giants to hunt for software vulnerabilities. Five months later, the program is no longer just a launch announcement: the New York Stock Exchange has told Congress it used the AI to fix flaws in its own infrastructure, a major application-security vendor has signed on, and Anthropic has shipped a materially stronger version of the restricted model at the center of it all.
None of that involves a newly announced $100 million commitment or a model called 'Mythos 2 Preview,' despite some recent write-ups describing exactly that. We looked for both and could not confirm either claim in Anthropic's own materials or in the outlets that broke the real September news — more on that further down. What we can verify is narrower, but still notable: a real customer testifying to real results, a real new partner, and a real model upgrade, all inside the same two-week window.
Quick summary — what's actually new since Project Glasswing's April 2026 launch: NYSE Group president Lynn Martin told the House Financial Services Committee on September 2, 2026 that NYSE used Project Glasswing to find and fix vulnerabilities in its own systems. Application-security vendor Checkmarx joined the program on September 3, 2026. Anthropic shipped Claude Mythos 5.1 — the restricted, safeguards-lifted sibling of Claude Fable 5.1 — on September 1, 2026, limited to vetted cybersecurity and life-sciences partners. Per Anthropic's own system card, Mythos 5.1 'substantially outperforms' Claude Opus 5 on cyber evaluations including ExploitBench, OSS-Fuzz, Firefox 147 and ExploitGym. We found no independent confirmation of a new $100M commitment or a 'Mythos 2 Preview' model this September — the only documented $100M figure traces back to the original April 2026 launch.
NYSE tells Congress the program is paying off
Speaking before the House Financial Services Committee on September 2, 2026, NYSE Group president Lynn Martin said the exchange had used Project Glasswing to identify and remediate weaknesses in its own systems, some of which had reportedly gone undetected for years. 'We have been able to find a variety of items that we've been able to very quickly address because of the state-of-the-art technology,' Martin said, according to Bloomberg's report on the testimony. She declined to detail the specific vulnerabilities or give a remediation timeline — a caution that's become typical among Glasswing partners, most of which treat their own findings as sensitive. NYSE is one of roughly 150 organizations now enrolled in the program, up from about 50 launch partners in April to about 150 total after the expansion announced in June 2026. Sources: [Bloomberg: NYSE Used Anthropic's Project Glasswing to Find Cyber Flaws](https://www.bloomberg.com/news/articles/2026-09-02/nyse-used-anthropic-s-project-glasswing-to-find-cyber-flaws), [CNBC: Anthropic expands Project Glasswing access to about 150 organizations](https://www.cnbc.com/2026/06/02/anthropic-mythos-ai-project-glasswing.html).
Checkmarx joins, citing a shrinking response window
A day later, on September 3, 2026, application-security firm Checkmarx announced it had joined Project Glasswing and gained access to Claude Mythos to strengthen its own vulnerability-detection work. 'We've watched frontier models surface risk that's gone undetected for years, faster than most organizations can keep up,' Checkmarx CEO Sandeep Johri said in the announcement. The company cited a J.P. Morgan analysis finding that roughly 80% of exploitations now occur on or before the day a vulnerability becomes public — a window that keeps shrinking as automated discovery tools, AI included, get faster. Checkmarx said it plans to publish its own findings on how AI is reshaping vulnerability triage and disclosure. Sources: [Checkmarx joins Anthropic's Project Glasswing](https://www.globenewswire.com/news-release/2026/09/03/3355880/0/en/checkmarx-joins-anthropic-s-project-glasswing.html).
The real upgrade: Claude Mythos 5.1
The more consequential September news is quieter. On September 1, 2026, Anthropic published the system card for Claude Fable 5.1 and Claude Mythos 5.1 — 'the same model, but with different levels of safeguards,' as Anthropic describes it. Fable 5.1 is the public-facing version, available through the API, Amazon Bedrock and Google Cloud Vertex AI at Fable 5's existing $10-per-million-input/$50-per-million-output-token pricing, with cache-read costs cut roughly 75% to $0.25 per million tokens. Mythos 5.1 is the restricted counterpart — the one actually deployed inside Project Glasswing — and Anthropic says it's 'available only through our trusted access programs,' with safeguards specifically shaped for partners working in cybersecurity and life sciences. Sources: [SD Times: Anthropic releases Claude Fable 5.1 and Mythos 5.1](https://sdtimes.com/claude-fable-5-1/61089/), [System Card: Claude Fable 5.1 & Claude Mythos 5.1](https://www-cdn.anthropic.com/0339e6a7c5c7b87f5c07798616dc32c215d14235/Claude%20Fable%205.1%20&%20Claude%20Mythos%205.1%20System%20Card.pdf).
According to Anthropic's own system card, Mythos 5.1 'substantially outperforms' Claude Opus 5 on nearly every cyber evaluation the company reports, including ExploitBench, OSS-Fuzz, Firefox 147 and ExploitGym — by Anthropic's account, the strongest vulnerability-hunting model it has released. That's a step up from Claude Mythos 5, which launched alongside the original Fable 5 on June 9, 2026, which was itself a step up from the Mythos Preview model that started the program in April. Anthropic hasn't published a fresh public benchmark for Mythos 5.1 comparable to the 73% expert-level capture-the-flag success rate the UK AI Security Institute measured for the original Preview model back in April — so the 'substantially outperforms' framing is currently Anthropic's own characterization, not yet an independent one.
About that '$100 million' and 'Mythos 2' claim
Some recent summaries circulating online describe this update as a new $100 million commitment tied to a model called 'Mythos 2 Preview.' We couldn't verify either detail. The only $100 million figure documented anywhere in Anthropic's materials is the one from Glasswing's original April 7, 2026 launch — '$100M in model usage credits' for partners, plus $4M in donations to open-source security organizations — and that figure keeps getting recirculated in coverage of later updates. As for naming, Anthropic's restricted cyber model has progressed Mythos Preview, then Mythos 5, then Mythos 5.1; nothing labeled 'Mythos 2' appears on Anthropic's own Glasswing page or in any of the outlets that covered the NYSE testimony, the Checkmarx announcement, or the Fable/Mythos 5.1 launch. If Anthropic announces a genuinely new, incremental financial commitment, it hadn't surfaced in the reporting we could find as of this writing. Sources: [Anthropic: Project Glasswing](https://www.anthropic.com/project/glasswing).
Project Glasswing by the numbers
| Milestone | April 2026 (Launch) | June 2026 (Expansion) | September 2026 (Now) |
|---|---|---|---|
| Restricted model | Claude Mythos Preview | Claude Mythos 5 (launched with Fable 5, June 9) | Claude Mythos 5.1 (launched Sept 1) |
| Partner organizations | ~50 launch partners | Expanded to ~150 total (from ~50) | ~150+, including newly confirmed NYSE and Checkmarx |
| Countries represented | Not disclosed at launch | 15+ | 15+ (unchanged) |
| Cumulative vulnerabilities found | Not yet quantified | 10,000+ high/critical severity | No newer public total confirmed |
| Headline cyber benchmark | 73% of expert-level CTF tasks (UK AISI) | 90.6% of 1,752 reviewed findings were true positives | "Substantially outperforms" Opus 5 on ExploitBench, OSS-Fuzz, Firefox 147, ExploitGym (Anthropic's system card) |
What decades-old bugs tell us
The recurring theme across every Glasswing update, including this one, is how old the bugs turn out to be. Earlier findings from the program's first cohort included:
- A vulnerability in OpenBSD that had persisted undetected for 27 years
- A remote-code-execution flaw in FreeBSD that went undetected for 17 years
- An FFmpeg bug that survived 16 years and more than five million automated fuzz tests
An independent review of 1,752 of the program's reported findings by six outside security firms validated 90.6% as genuine true positives, with 62.4% confirmed high or critical severity — an accuracy figure that matters as much as the raw discovery count. NYSE's and Checkmarx's September announcements fit the same pattern: long-buried issues, found fast, by a model whose own creators say they still can't fully explain how it chains multiple weaknesses into a working exploit path. Sources: [CSO Online: Project Glasswing has uncovered 10,000 vulnerabilities](https://www.csoonline.com/article/4176865/project-glasswing-has-uncovered-10000-vulnerabilities-anthropic.html).
Every Mythos release so far — Preview, 5, and now 5.1 — has shipped alongside a publicly downloadable sibling (Fable) with the offense-enabling safeguards put back in. If you want a sense of what the underlying model can do without vetted-partner access, Fable 5.1 is as close as the public gets.
01Did Anthropic really commit a new $100 million to Project Glasswing in September 2026?
We couldn't confirm that. The only documented $100 million commitment is the one from Glasswing's original April 7, 2026 launch, covering model usage credits for partners plus $4 million in donations to open-source security organizations. No Anthropic material or news outlet we checked describes an additional September pledge.
02What is Claude Mythos 5.1?
It's the restricted, safeguards-lifted version of Claude Fable 5.1, Anthropic's model pair that launched September 1, 2026. Per Anthropic's system card, it posts the company's strongest cyber-evaluation scores to date, but access stays limited to vetted Project Glasswing partners in cybersecurity and life sciences — it isn't sold or generally available.
03Which companies joined Project Glasswing most recently?
NYSE (confirmed via September 2, 2026 Congressional testimony) and application-security vendor Checkmarx (announced September 3, 2026) are the most recently confirmed additions, joining a roster that's grown from about 50 launch partners in April to roughly 150 organizations by mid-2026.
04Can regular users try Claude Mythos?
No. Claude Mythos has never been made publicly available in any version — Preview, 5, or 5.1 — and Anthropic has said uncontrolled release could make large-scale cyberattacks more likely. The public equivalent is Claude Fable, currently at version 5.1.
Claude Mythos stays behind Anthropic's vetted-partner wall, but the public Claude models it ships alongside — Fable 5.1 included — are available to anyone. LumiChats lets you compare Claude against ChatGPT, Gemini, Grok and other leading chatbots side by side, or chat with several of them in one place, so you can judge the public models for yourself without juggling separate subscriptions.
