Anthropic Accidentally Leaked Claude Mythos (Everything We Know)
Speakers
Summary
- The accidental leak of Anthropic's advanced AI model, Claude Mythos (within a new 'Capybara' tier), reveals a "step change" in AI performance, dramatically outperforming Claude Opus 4.6, especially in coding and cybersecurity benchmarks.
- The model's power poses a significant security threat; its ability to discover and exploit vulnerabilities is a primary reason for its controlled release, initially only to cybersecurity experts to harden defenses.
- The market reacted negatively to the leak, with cybersecurity stocks like CrowdStrike and Palo Alto Networks losing billions, echoing a pattern of dips following AI security announcements.
- The intense compute required to train and run Mythos is likely causing service degradation for current Claude users, highlighting a growing infrastructure strain as labs pursue larger models.
- The competitive landscape is intensifying: OpenAI is working on a rival model (codename "Spud"), focusing resources after a period of distraction, while Google is developing "Agent Smith" to compete in AI coding.
- Internal signals suggest these are "step function improvements" with novel capabilities; an OpenAI internal memo described progress moving "faster than many of us expected."
- Prediction markets (Polymarket) now heavily favor Anthropic (66% chance) to have the best AI model by June, a significant shift from Google being the favorite in February.
- The race is creating tension and trade-offs, with labs potentially sacrificing current service quality and facing financial/logistical challenges (e.g., OpenAI defaulting on GPU/data center deals) to fund the next breakthrough.
Speakers
Summary
- Anthropic accidentally leaked Claude Mythos, a model with step-change performance above Opus 4.6, particularly in cybersecurity, leading to controlled release to experts.
- The leak resulted from a CMS configuration error, exposing nearly 3,000 unpublished assets including the new Capybara tier and Mythos model details.
- Claude Mythos dramatically outperforms Opus 4.6 in coding, academic reasoning, and cybersecurity benchmarks, with capabilities so advanced it poses dual-use risks.
- Market reaction included dips in cybersecurity stocks like CrowdStrike and Palo Alto Networks, with similar drops occurring monthly on AI security news.
- Josh questions if immediate market concern is warranted, citing the model's compute intensity and deployment challenges as limiting factors.
- AI progress is accelerating, with competitors OpenAI (codenamed Spud) and Google (Agent Smith) developing larger models (10-20 trillion parameters) for near-term release.
- OpenAI is refocusing by killing projects like Sora and reallocating resources to core models, amid financial strains from defaulted GPU and data center deals.
- Compute constraints are affecting current AI services, causing Claude server outages and reduced user experience due to resource allocation for training.
- Polymarket predictions show Anthropic with a 66% chance of having the best AI model by June, a significant shift from Google's earlier lead.
- The competitive landscape is tense among Anthropic, OpenAI, and Google, with each pushing for breakthroughs ahead of potential IPOs and economic impact.
- Key uncertainties include specific model capabilities, release timelines, and the broader implications for cybersecurity, compute infrastructure, and stock valuations.
Trade Ideas
The speaker explicitly stated that CrowdStrike and Palo Alto Networks stock prices dropped ("were down a couple billion") on the news of Anthropic's Claude Mythos leak and its advanced cybersecurity capabilities. Advanced AI models like Claude Mythos represent an existential technological threat to traditional cybersecurity firms by automating the discovery and exploitation of vulnerabilities at a superhuman scale and speed, potentially disrupting their core value proposition. The market's negative reaction and the recurring pattern ("happening seemingly on a monthly basis") suggest these stocks are vulnerable to de-risking as AI capabilities advance, making them an area to avoid in the near term. The immediate commercial rollout of such powerful, compute-intensive AI models may be slow, giving incumbents time to adapt or integrate the technology themselves.
Up Next
This Bankless video, published March 31, 2026, features Ejaaz Ahamadeen discussing PANW, CRWD. 1 trade idea extracted by AI with direction and confidence scoring.
Speakers: Ejaaz Ahamadeen · Tickers: PANW, CRWD