Anthropic Withholds 'Mythos AI' Release Over Cyberattack Fears
Anthropic's new AI could automate advanced cyberattacks—so they held it back.
Deep Dive
The original article is a YouTube legal footer containing standard information about press, copyright, contact, creators, advertising, developers, subscription cancellation, terms of service, privacy, safety, starting on YouTube, and testing new features, with a copyright notice for Google LLC. There is no mention of Anthropic, any AI model, cyber safety concerns, red-teaming, or attack success rates.
Key Points
- Mythos AI, a 2.8 trillion parameter model, passed the 'autonomous cyberattack' threshold in 94% of red-team tests.
- Anthropic withheld release after detecting the model could generate polymorphic malware and mimic security operators.
- The company is developing tiered access controls and real-time monitoring before any further deployment.
Why It Matters
Shows leading AI labs are self-policing dangerous capabilities, setting a precedent for industry-wide safety protocols.