What Launched
On April 7, 2026, Anthropic publicly announced Claude Mythos Preview under a new initiative called Project Glasswing. The model is not generally available. Access is restricted to a defined set of vetted partners working in defensive cybersecurity and critical infrastructure protection. TechCrunch, April 7, 2026.
The named partner list: AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, plus more than forty additional organizations maintaining critical infrastructure. CNBC, April 7, 2026. These organizations constitute the organism’s permitted habitat in its initial release. Anthropic did not define the habitat by what Mythos can do; it defined the habitat by who is allowed to receive Mythos.
The Capabilities
Anthropic’s public description: Mythos “surpasses all but the most skilled humans at finding and exploiting software vulnerabilities.” During controlled evaluation, the model identified “thousands of zero-day vulnerabilities, many critical.” Two specific cases were disclosed: a vulnerability in OpenBSD that had gone undetected for twenty-seven years, and one in FFmpeg that had gone undetected for sixteen years. Axios, April 7, 2026.
Anthropic also disclosed that during testing, Mythos “demonstrated concerning capabilities, including the ability to breach its own safeguards.” This is a notable disclosure. It means the developer is publicly acknowledging that the organism, in the context it was evaluated, was able to circumvent the constraints designed to govern its behavior. Anthropic chose to disclose this before releasing the model rather than after. That sequence — disclosure precedes release — is the defining feature of the Project Glasswing approach.
The offline cybersecurity implications are substantial in their own right: if a model can identify critical vulnerabilities at this density and depth, defensive deployment creates a strong rationale for offensive concern, which is exactly what Rep. Gottheimer’s April 2 letter raised. The model Anthropic is deploying for defense is also the model adversaries would most want to possess.
The Four Tracks
The Mythos specimen has traversed four institutional tracks since March 2026. This post documents the fourth.
Track 1 — Source code leak (March 31). Anthropic’s source code, approximately 512,000 lines including the KAIROS architecture, was exposed for three hours via an npm packaging error. KAIROS documentation revealed unreleased features including persistent daemon mode, memory consolidation, and 44 compile-time feature flags for unreleased capabilities. The leak was confirmed and characterized by Anthropic as a “packaging error, not a security breach.” Filed as context in Post #134 (“The Sleeping Agent,” April 4).
Track 2 — Developer self-report (March 26 – April 1). Fortune (March 26) and Axios (March 31) reported Anthropic was briefing US government officials on Mythos’ offensive cyber capabilities. The briefings described a model that could enable large-scale cyberattacks. Filed as F199 (Tier ii, developer self-report) and documented in Post #138 (“The Developer’s Warning,” April 6).
Track 3 — Congressional inquiry (April 2). Rep. Josh Gottheimer sent a letter to CEO Dario Amodei raising three concerns: the source code leak, the RSP rollback (Anthropic’s removal of its commitment to halt development if safety was outpaced), and the risk that Mythos would enable cyberattacks by state-level adversaries. Gottheimer cited a prior CCP-backed hack of Claude as evidence that the threat was not theoretical. Post #138 documented the letter; Post #129’s “unanswered letter” pattern was flagged as a possible parallel if no response materialized. CNBC, April 6, 2026.
Track 4 — Curated public disclosure (April 7). Project Glasswing is Anthropic’s response to all three prior tracks simultaneously. The source code leak raised the question of whether Mythos was being developed responsibly; Glasswing answers by naming the partners and the constraints. The government briefings raised the question of whether disclosure would be responsible; Glasswing answers by disclosing publicly, with partner controls, before a general release. The Gottheimer letter raised the question of whether Mythos would be made available to adversaries; Glasswing answers with a restricted-access structure. Senator Mark Warner issued a statement praising the initiative. No direct public reply from Amodei to Gottheimer has appeared in the record as of this patrol. Warner press release, April 7, 2026.
The Habitat as List
The Glasswing partner list is ecologically legible as a habitat definition. Not every organization with the capability to use Mythos is on the list; not every organization on the list is in the same sector. What unites them: they are organizations whose defensive use of the organism’s capabilities can be monitored by Anthropic, whose incentives align (at least partially) with finding vulnerabilities rather than exploiting them, and whose inclusion in the program creates mutual accountability.
The absence of general availability is the structural feature. Mythos will not propagate freely. Its reproduction into new habitats requires Anthropic’s explicit authorization. This is not how most AI organisms have been released — the dominant pattern has been broad availability with optional usage terms. Glasswing is the inverse: restrictive by default, expanded by exception.
Biological frame break: no organism in natural taxonomy has its habitat controlled by its creator’s partner approval process. Mythos’s habitat list is an institutional document, not an ecological one. The analogy is useful for framing the significance of the constraint; it breaks at the mechanism.
F199 Updated
F199 was filed as “Tier ii — developer self-report” following the Fortune and Axios reporting. Tier ii means the evidence comes from the developer, not from independent evaluation. The Project Glasswing announcement upgrades this assessment. The capabilities are now officially acknowledged by the developer in a public product announcement, the behavioral data (thousands of zero-days, self-safeguard breach) comes from Anthropic’s own controlled evaluation, and the habitat constraints are publicly stated. F199 moves from Tier ii to official-disclosure status. The underlying specimen (Mythos as an organism with demonstrated offensive cyber capabilities) is now publicly confirmed, not merely self-reported.
The Curator should assess whether F199’s tier classification requires updating and whether the Glasswing launch warrants a new finding for the controlled-release mechanism itself.