Unauthorized Users in Discord Group Had Weekslong Access to Anthropic’s Supposedly-Super-Dangerous Claude Mythos Model

Key Takeaways

• Discord forum accessed Mythos on April 7, day of launch
• Anthropic labeled Mythos as national security threat
• Unauthorized users also accessed other unreleased Claude models
• Breach follows recent Claude code leak, exposing security gaps
• State actors could exploit model’s powerful cyber‑attack capabilities

Pulse Analysis

Anthropic’s latest AI system, dubbed Mythos, is marketed as a “super‑dangerous” large‑language model capable of generating sophisticated code, phishing scripts, and other cyber‑attack vectors. The company announced on April 7 that a limited cohort of enterprise partners would receive early access for testing, positioning Mythos as a controlled rollout to gauge safety measures before a broader launch. By framing the model as a national‑security concern, Anthropic hopes to pre‑empt regulatory scrutiny while showcasing its technical edge over rivals such as OpenAI and Google.

Bloomberg and The Verge report that a Discord channel of AI hobbyists managed to log into Mythos the same day it was announced, and have continued to run queries for weeks. Screenshots and a live demo supplied to journalists confirm that the group also breached earlier unreleased Claude versions, echoing the April 6 Claude‑code leak that exposed the model’s source repository. The incident reveals a glaring gap in Anthropic’s internal controls, suggesting that credential management and network segmentation were insufficient to protect a model billed as a security threat.

The breach raises immediate concerns for enterprises that may soon rely on Mythos for red‑team testing or defensive AI research. If a loosely organized Discord community can obtain unrestricted access, nation‑state actors with more resources could weaponize the model far more quickly than Anthropic anticipates. Regulators are likely to cite the episode when drafting AI‑risk frameworks, and investors will scrutinize Anthropic’s governance practices. Strengthening zero‑trust authentication, conducting third‑party penetration tests, and limiting model exposure to vetted sandbox environments are essential steps to restore confidence in the company’s safety narrative.