Anthropic Mythos Model Leaked via Contractor and Guesswork
Why It Matters
This breach undermines the narrative that frontier AI labs can safely sequester models with existential risks. It highlights the extreme vulnerability of AI infrastructure to social engineering and supply chain weaknesses.
Key Points
- Unauthorized users accessed Anthropic's Mythos model on its announcement day through a private Discord group.
- The breach was facilitated by an Anthropic third-party contractor and data previously leaked from AI startup Mercor.
- Anthropic had explicitly withheld Mythos from the public, citing it as too dangerous for general release.
- The group reportedly maintains active access to the model and has been using it continuously since the breach.
- The incident underscores the fragility of AI safety protocols against social engineering and insider threats.
Anthropic’s highly guarded Mythos AI model has reportedly been compromised by unauthorized users who successfully located the program on the day of its public announcement. According to reports from Bloomberg and Fortune, a private Discord group gained access by leveraging information provided by an Anthropic third-party contractor and data previously stolen from the startup Mercor. While Anthropic had categorized Mythos as too dangerous for public release due to its advanced capabilities, the group has reportedly been running the model continuously without intervention. The incident highlights significant security gaps in the safeguarding of frontier AI systems, particularly regarding the human element of third-party contractors and the predictability of internal infrastructure. Anthropic has not yet confirmed if they have regained exclusive control over the model's environment or if the weights have been exfiltrated entirely.
Anthropic’s secret, super-dangerous model called Mythos just got leaked in a way that feels like a heist movie. A small group on Discord basically guessed where the model was hidden after getting tips from a contractor working for Anthropic. It is as if someone left a vault unlocked because they used a predictable password and a friend on the inside gave a hint. Even though Anthropic said the world was not ready for this AI, these users have been playing with it ever since. This is a massive wake-up call for the industry.
Sides
Critics
Unauthorized users who exploited insider knowledge to prove that the model's security was insufficient.
Defenders
Claims the model is too dangerous for public release and is now managing a significant security breach.
Neutral
An AI training startup whose previous data leak allegedly provided the roadmap for locating Mythos.
Noise Level
Forecast
Anthropic will likely face intense regulatory scrutiny regarding their internal security protocols and third-party vetting processes. Expect a shift in the industry toward air-gapped training environments for frontier models to prevent similar credential-based leaks.
Based on current signals. Events may develop differently.
Timeline
Media Reports Leak
Bloomberg and Fortune report that the group is still actively using the model.
Unauthorized Access Achieved
A private Discord group uses contractor knowledge and previous leaks to guess the model's location and gain access.
Mythos Announced
Anthropic publicly announces the Mythos model but states it will remain private due to safety concerns.
Mercor Data Leak
Information regarding Anthropic's internal practices is stolen from the startup Mercor by hackers.
Join the Discussion
Discuss this story
Community comments coming in a future update
Be the first to share your perspective. Subscribe to comment.