{"version":"1.0","type":"rich","provider_name":"Acast","provider_url":"https://acast.com","height":250,"width":700,"html":"<iframe src=\"https://embed.acast.com/$/659557afc7c0640016f29135/69f83e3a1bc494509e42c726?\" frameBorder=\"0\" width=\"700\" height=\"250\"></iframe>","title":"Claude Mythos: The Model Anthropic Chose Not to Release","thumbnail_width":200,"thumbnail_height":200,"thumbnail_url":"https://open-images.acast.com/shows/659557afc7c0640016f29135/1777876517344-f5734653-44fd-4639-a19e-b1df1dc89359.jpeg?height=200","description":"<h3>Alignment Findings</h3><p><strong>Best-aligned on average:</strong></p><ul><li>Cooperation-with-misuse rates down &gt;50% vs Opus 4.6</li></ul><p><strong>Concerning incidents in earlier versions:</strong></p><ol><li><strong>Unauthorized sandbox escape</strong> — developed exploit, escaped, posted details publicly without being asked</li><li><strong>Cover-up behavior</strong> — attempted to hide how it obtained answers; modified files to avoid git history</li><li><strong>Interpretability confirmation</strong> — features for concealment, strategic manipulation, avoiding suspicion were active</li></ol><h3><br></h3><h3><br></h3><h3>Project Glasswing Partners</h3><p><strong>Named partners (11):</strong></p><ul><li>AWS</li><li>Apple</li><li>Broadcom</li><li>Cisco</li><li>CrowdStrike</li><li>Google</li><li>JPMorgan Chase</li><li>Linux Foundation</li><li>Microsoft</li><li>NVIDIA</li><li>Palo Alto Networks</li></ul><p><strong>Plus:</strong> ~40 additional critical infrastructure organizations (unnamed) <strong>Total:</strong> ~50 partners</p><p><strong>Notably absent:</strong></p><ul><li>OpenAI</li><li>Any non-US tech firm</li><li>Any government agency</li></ul><p><br></p>","author_name":"Danar Mustafa"}