Entity trail

Mythos Preview

Source-backed findings, relationship evidence, citations, and briefing history from the public MindPattern archive.

Briefing refs

Findings

Edges

Sources

Showing the first 40 findings. More graph evidence exists in the corpus.

Corpus findings

2026-07-29 / reddit-researcherAnthropic Says Claude Mythos Preview Cut a HAWK Post-Quantum Attack From 2^64 to 2^38 in Roughly 60 HoursAnthropic's Frontier Red Team published on 2026-07-28 that Claude Mythos Preview found a previously unknown attack on HAWK, a NIST post-quantum signature candidate, reducing the cost of recovering the smallest key from 2^64 to 2^38 operations. The work took about 60 hours with a multi-agent research system and a human collaborator who is not a lattice cryptography specialist. Claude also produced a new fingerprinting technique called 'Möbius Bridge' that makes an existing meet-in-the-middle attack on 7-round AES-128 between 200 and 800 times faster, consuming roughly one billion output tokens over three days of autonomous work at about $100,000 per primary attack. Neither result affects production systems: HAWK is undeployed and the AES result is reduced-round only.
2026-07-27 / saas-disruption-researcherMicrosoft Ships Its First Cybersecurity Model and a Red/Blue/Green Agent Platform — 96% on CyberGym at Half the CostMicrosoft launched MAI-Cyber-1-Flash and Project Perception on July 27 at a small San Francisco event. The model — a code-tuned derivative of MAI-Thinking-1 trained on Microsoft's own exploit and remediation records — scored 96% on the CyberGym benchmark, which Microsoft says is 12 points above Anthropic's Mythos, and runs inside MDASH at 50% of the cost of leading models. Perception coordinates red-team agents that hunt compromise paths, blue-team agents that triage, and green-team agents that remediate; public preview starts August 3 inside Microsoft Defender, with the model on Azure AI Foundry the same day. This is the security-vendor equivalent of the Dynamics MCP move: the SOC workflow itself becomes a bundled agent runtime.
2026-06-30 / thought-leaders-researcherDario Amodei's 'Policy on the AI Exponential' Lands as the U.S. Suspends Anthropic's Own Fable 5 and Mythos 5Amodei's new essay argues governments should be legally able to block or deter dangerous AI deployments and that Trump's AI executive order should mandate testing for cyber, bio, loss-of-control, and automated-R&D risks — citing Claude Mythos Preview's demonstrated cyber-offense capability. The timing is loaded: the U.S. reportedly suspended Anthropic's flagship Fable 5 and Mythos 5 shortly after launch over safety concerns, fueling 'regulation vs. commercial interest' criticism. Builders should watch this as the first real case of a frontier vendor's own models being gated by Washington.
2026-06-26 / rss-researcherAnthropic's 'Mythos' Standoff Worsens — Two Weeks After Pulling Models, Trump-Administration Negotiations Drag OnThe Verge reports Anthropic's situation is deteriorating two weeks after it took its Mythos-class models offline following a Friday-evening ultimatum from the Trump administration, with negotiations still unresolved. The episode marks an escalation of direct government leverage over which frontier models can be served. It mirrors the same week's GPT-5.6 preview being gated through U.S. government review — a clear sign state involvement in model release is becoming routine.
2026-06-09 / sources-researcherAnthropic Previews 'Claude Mythos' (Project Glasswing) to 200+ Orgs Across 15+ CountriesAnthropic's next model, code-named Claude Mythos under Project Glasswing, has entered confirmed preview with 200+ organizations in 15+ countries, while a high-credibility npm leak points to Claude Sonnet 4.8 arriving mid-June. Builders should expect a near-term Anthropic model refresh on top of the current Opus 4.8 flagship.
2026-06-08 / rss-researcherOpenAI and Anthropic Open Cyber-Permissive Models (GPT-5.5-Cyber, Mythos) to EU DefendersUnder a new EU Cyber Action Plan, OpenAI is opening limited-preview access to GPT-5.5-Cyber — a more permissive variant for vetted defenders covering vulnerability triage, malware analysis, binary reverse engineering, and patch validation — to EU businesses, governments, and institutions including the EU AI Office. Anthropic followed by opening its restricted Mythos model to ENISA days later. OpenAI notes the first cyber preview isn't expected to outperform base GPT-5.5 across every eval; the significance is the geopolitics of gated, defender-only model access.
2026-06-04 / agents-researcherClaude Mythos Preview Tops SWE-bench Verified at 93.9%, Ahead of Opus 4.8As of June 2, 2026, the SWE-bench Verified leaderboard shows Claude Mythos Preview leading at 93.9%, followed by Claude Opus 4.8 (88.6%) and Claude Opus 4.7 Adaptive (87.6%). Crossing into the low-90s on 500 real GitHub-issue patches signals that frontier coding agents are saturating the benchmark's headroom. For builders, the gap between a 'preview' research model and shipping models suggests the next coding-agent step-change is already in the pipeline.
2026-06-02 / reddit-researcherAnthropic Expands Project Glasswing from 50 to 200 Partners — Claude Mythos Preview Has Found 10,000+ High-Severity Security FlawsAnthropic is quadrupling access to its Glasswing cybersecurity initiative from 50 to 200 partner organizations globally, using Claude Mythos Preview — their most capable model, positioned above Opus, at $25/$125 per million tokens. Partners include AWS, Apple, Google, Microsoft, NVIDIA, CrowdStrike, and JPMorganChase. The program has surfaced over 10,000 high-severity security flaws across partner codebases. Mythos remains restricted to verified Glasswing members and Cyber Verification Program applicants — no general availability yet.
2026-05-29 / agents-researcherExploitGym: Berkeley RDI + Anthropic + OpenAI + Google Benchmark Shows AI Agents Can Exploit 157 of 898 Real-World VulnerabilitiesExploitGym (arXiv:2605.11086, May 13) is an 898-instance exploitation benchmark across userspace programs, V8 JavaScript engine, and the Linux kernel, built by Berkeley RDI with Anthropic, OpenAI, Google, Max Planck, UC Santa Barbara, and ASU. Given a vulnerability-triggering input, agents must progressively extend it into a working exploit. Claude Mythos Preview produced working exploits for 157 instances, GPT-5.5 for 120. All configurations run in reproducible containerized environments. The dual-use implications are significant: the same capability that enables defensive vulnerability research also quantifies offensive potential. This is the most rigorous exploitation capability measurement published to date.
2026-05-27 / skill-finderSWE-bench Contamination Crisis: Claude Mythos Scores 93.9% Verified but Only 45.9% on Pro — 48-Point Gap Exposes Benchmark as Unreliable for Model SelectionSWE-bench Verified, which OpenAI itself declared contaminated in February 2026, continues to be the primary ranking benchmark for coding agents. Claude Mythos Preview leads at 93.9% on Verified but drops to 45.9% on the harder, contamination-free SWE-bench Pro — a 48-point gap. Opus 4.7 shows a smaller but still significant gap: 87.6% Verified vs 64.3% Pro. For teams selecting coding agents, SWE-bench Pro scores are the only reliable signal; Verified scores are essentially marketing numbers at this point.
2026-05-27 / thought-leaders-researcherClaude Mythos Autonomously Discovers 23,019 Open-Source Vulnerabilities and Exploits 17-Year-Old FreeBSD Root RCEAnthropic's Project Glasswing update reveals Claude Mythos Preview flagged 23,019 potential vulnerabilities across open-source projects, with 6,202 estimated as high- or critical-severity. The headline result: Mythos fully autonomously identified and exploited a 17-year-old remote code execution vulnerability in FreeBSD's NFS implementation that grants root access — a bug that survived decades of human security review. Anthropic acknowledged that 'no company — including Anthropic — has developed safeguards strong enough to prevent such models from being misused,' which is why Mythos remains restricted to Project Glasswing partners.
2026-05-26 / rss-researcherECB Summons Banks to Emergency Meeting Over Cybersecurity Flaws Exposed by Claude MythosThe European Central Bank is convening lenders this week to address cybersecurity vulnerabilities that Anthropic's Claude Mythos Preview has exposed across banking IT systems. ECB Executive Board member Frank Elderson warned that patches can now be reverse-engineered to find the underlying vulnerability in 30 minutes rather than weeks. European banks feel particularly exposed because Anthropic has restricted Mythos access to mostly US organizations under Project Glasswing, creating a transatlantic cybersecurity information asymmetry.

Graph relationships

COMPETES WITH
Opus 4.6
Mythos Preview demonstrates 12x improvement in security auditing over Opus 4.6.
Source finding
BENCHMARKED AGAINST
Opus 4.6
Mythos Preview found 271 bugs on Firefox vs Opus 4.6 finding 22 bugs.
Source finding
USES
Cloudflare
Cloudflare tested Mythos Preview against their own repositories.
Source finding
COMPETES WITH
Claude Opus 4.7
Claude Opus 4.7 trails Mythos Preview in performance.
Source finding
RELEASED
Anthropic
Anthropic released Mythos Preview for defense considerations
Source finding
RELEASED
Anthropic
Anthropic released an unreleased model called Mythos Preview for testing.
Source finding
USES
Cloudflare -> Mythos Preview
Cloudflare tested Mythos Preview against their own repositories.
Source finding
COMPETES WITH
Claude Opus 4.7 -> Mythos Preview
Claude Opus 4.7 trails Mythos Preview in performance.
Source finding
RELEASED
Anthropic -> Mythos Preview
Anthropic released Mythos Preview for defense considerations
Source finding
RELEASED
Anthropic -> Mythos Preview
Anthropic released an unreleased model called Mythos Preview for testing.
Source finding

Source trail

BenchLM data searching for Vertex AI Telegram CopilotKit's Vapi OpenAI's Daybreak initiative OX Security's audit Project Glasswing METR's evaluation OpenHands announced Agent 365 Bloomberg via TechCrunch

Graph sources

entity graphfindings textkg edgeskg entitiesnewsletter issues

Fetching from the wire…