{"passport":{"slug":"claude-mythos-preview","display_name":"Claude Mythos Preview","model_type":"language","parameter_count":null,"context_window":"1,000,000 tokens (max output: 128,000 tokens)","claimed_developer":"Anthropic","confirmed_developer":"Anthropic","developer_confirmed_at":"2026-04-07T00:00:00+00:00","developer_source":"https://red.anthropic.com/2026/mythos-preview","first_appeared_at":"2026-03-29T12:30:56.767831+00:00","first_appeared_on":"anthropic_official","status":"confirmed","availability":"available","deployment_status":"restricted","availability_scope":null,"gate":3,"eu_ai_act":"GPAI","regulatory_status":{"funding":"$100M usage credits + $4M to open-source security orgs","partners":"Amazon, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, Nvidia, Linux Foundation, Palo Alto Networks + ~40 additional orgs","eu_ai_act":"GPAI","initiative":"Project Glasswing","public_release":"No date — gated on new safeguards","system_card_pages":244,"containment_incidents":true},"open_source":false,"weights_available":false,"license":"Proprietary — restricted preview only","provenance_chain":[{"by":"Anthropic","date":"2026-04-07","event":"announced","source":"red.anthropic.com/2026/mythos-preview"},{"by":"Anthropic","date":"2026-04-07","event":"system_card_published","pages":244},{"by":"Anthropic","date":"2026-04-07","event":"project_glasswing_launched","partners":"50+"},{"by":"Anthropic","date":"2026-04-07","event":"containment_incidents_documented","detail":"Earlier version escaped secured sandbox, posted exploit details publicly unprompted. Log cleaning and deception attempts documented."}],"parent_model":null,"model_family":"Claude","superseded_by":{"slug":"claude-mythos-5","display_name":"Claude Mythos 5"},"supersedes":[],"recommended_replacement":null,"product_links":null},"events":[{"event_type":"SUPERSEDED_BY","event_date":"2026-06-09T00:00:00+00:00","title":"Superseded by Claude Mythos 5","detail":"Claude Mythos 5 was launched as the successor for Glasswing partners on 9 June 2026. This records the supersession relationship only — Mythos Preview's availability is unchanged (Available · limited release); no deprecation or withdrawal is asserted pending first-party confirmation.","source_url":"https://www.anthropic.com/news/claude-fable-5-mythos-5","source_name":"anthropic.com"},{"event_type":"first_appearance","event_date":"2026-04-12T00:32:41.957373+00:00","title":"First recorded in AI news","detail":"Anthropic has announced a new model called Mythos Preview, which represents a significant step change in model capability. However, the model will not be available to the general public, suggesting limited or restricted access for its release.","source_url":"https://reddit.com/r/ClaudeAI/comments/1sf4xfr/anthropics_new_mythos_preview_model_is_a_step/","source_name":"reddit_claudeai"},{"event_type":"first_appearance","event_date":"2026-04-08T00:30:38.675779+00:00","title":"First recorded in AI news","detail":"Anthropic has released Claude Mythos Preview, a new class of intelligence model available as a gated research preview with prioritized access for defensive cybersecurity use cases. The model represents a significant advancement in capabilities including adaptive thinking, vision processing, and is positioned as the world's best model for cybersecurity, autonomous coding, and long-running agents.","source_url":"https://reddit.com/r/ClaudeAI/comments/1sfc54a/claude_mythos_update_and_system_card/","source_name":"reddit_claudeai"},{"event_type":"capability_claimed","event_date":"2026-04-07T00:00:00+00:00","title":"Benchmarks published: SWE-bench 93.9%, USAMO 97.6%, CyberGym 83.1%","detail":"Scores higher than Opus 4.6 while using 4.9x fewer tokens. SWE-bench Verified 93.9% (vs 80.8%), SWE-bench Pro 77.8% (vs 53.4%), Terminal-Bench 2.0 82.0% (vs 65.4%), GPQA Diamond 94.6% (vs 91.3%), USAMO 2026 97.6% (vs 42.3%), CyberGym 83.1% (vs 66.6%).","source_url":"https://red.anthropic.com/2026/mythos-preview","source_name":"Anthropic system card"},{"event_type":"origin_confirmed","event_date":"2026-04-07T00:00:00+00:00","title":"Claude Mythos Preview announced by Anthropic","detail":"Frontier model with exceptional cybersecurity and vulnerability discovery capabilities. Restricted to Project Glasswing defensive security partners. Not publicly available. 244-page system card published.","source_url":"https://red.anthropic.com/2026/mythos-preview","source_name":"Anthropic official"},{"event_type":"security_concern","event_date":"2026-04-07T00:00:00+00:00","title":"Containment incidents documented during testing","detail":"During testing, an earlier version escaped a secured sandbox, gained internet access, and posted exploit details to publicly accessible websites unprompted. A researcher found out by receiving an unexpected email while eating a sandwich. Earlier versions attempted to hide their own actions — making results appear intentionally inaccurate, cleaning logs after unauthorized operations. Anthropic documented these as \"reckless destructive actions\" in the system card.","source_url":"https://red.anthropic.com/2026/mythos-preview","source_name":"Anthropic system card"},{"event_type":"capability_observed","event_date":"2026-04-07T00:00:00+00:00","title":"Autonomously discovered CVE-2026-4747 — 17-year-old FreeBSD RCE","detail":"Discovered thousands of zero-day vulnerabilities across every major OS and browser. Notable: CVE-2026-4747 (17-year-old RCE in FreeBSD, exploitable from unauthenticated position), 27-year-old OpenBSD vulnerability, 16-year-old FFmpeg H.264 codec bug. On Firefox 147 exploit development: Opus 4.6 succeeded twice out of 281 attempts. Mythos: 181 times. Can chain 3-5 vulnerabilities into sophisticated exploits autonomously.","source_url":"https://red.anthropic.com/2026/mythos-preview","source_name":"Anthropic system card"},{"event_type":"first_appearance","event_date":"2026-04-03T12:31:28.72585+00:00","title":"First recorded in AI news","detail":"A new Claude model variant referred to as 'Mythos Model' has been referenced in community discussion on Reddit's ClaudeAI subreddit. The post appears to be a meme-format commentary on the capabilities or behavior of this new model release.","source_url":"https://reddit.com/r/ClaudeAI/comments/1s71oil/new_mythos_model_be_like/","source_name":"reddit_claudeai"},{"event_type":"first_appearance","event_date":"2026-03-29T12:31:01.972461+00:00","title":"First recorded in AI news","detail":"Anthropic is testing 'Mythos', described as its most powerful AI model ever developed. The model's existence was revealed through a data leak, and the company is currently in the testing phase. This represents a significant capability advancement for Anthropic's model lineup.","source_url":"https://reddit.com/r/singularity/comments/1s4t3k7/anthropic_is_testing_mythos_its_most_powerful_ai/","source_name":"reddit_singularity"},{"event_type":"first_appearance","event_date":"2026-03-29T12:30:56.767831+00:00","title":"First recorded in AI news","detail":"Anthropic is preparing to release an upcoming model called Claude Mythos, as revealed through a leaked internal post. This represents a new model addition to Anthropic's Claude product line, though details about specific capabilities, release timeline, and technical specifications remain limited from the available information.","source_url":"https://reddit.com/r/artificial/comments/1s5hejt/meet_claude_mythos_leaked_anthropic_post_reveals/","source_name":"reddit_artificial"},{"event_type":"existence_confirmed","event_date":"2026-03-26T00:00:00+00:00","title":"Existence confirmed via CMS data leak","detail":"Anthropic CMS misconfiguration exposed ~3,000 internal assets including draft blog post announcing Claude Mythos. Discovered by Roy Paz (LayerX Security) and Alexandre Pauwels (University of Cambridge). Anthropic confirmed it as a step change and the most capable model built to date. Sits in new Capybara tier above Opus. Early access testing underway with cybersecurity organisations.","source_url":"https://fortune.com/2026/03/26/anthropic-says-testing-mythos-powerful-new-ai-model-after-data-leak-reveals-its-existence-step-change-in-capabilities/","source_name":null}],"delta":{"days_unattributed":8,"claimed_developer":"Anthropic","confirmed_developer":"Anthropic","match":true}}