Mythos preview pulls the industry back to defense: The fundamental security of DeFi is shaken, and AI resources are flooding into the risk control track.

Mythos Shifts AI Industry Toward Defense: Several Key Issues Surface

Deedy Das’s tweet not only sparked discussion but also shifted the entire industry from “how powerful the model is” back to “what strong models can do and how to regulate them.” According to publicly available information, Mythos Preview scored 99/100 on BenchLM, full marks on proxy tasks, and approximately 94% on SWE-bench coding benchmarks, significantly higher than Claude Opus 4.6. Reports indicate that this model used less than $50 worth of compute power to locate a vulnerability from 27 years ago in OpenBSD and can chain Linux kernel vulnerabilities on its own. Karpathy, LeCun, and others have yet to publicly comment, possibly because the model is not yet open to independent researchers.

My judgment: Industry attention is shifting from “capability rankings” to “defense and regulation.” Anthropic is managing limited access through Project Glasswing, aiming to address dual-use issues before wider deployment.

• Enterprise access creates resource and information gaps: Glasswing has already onboarded over 40 partners (including Apple and Microsoft), with a $100 million security scenario fund; open-source projects find it hard to access comparable resources.
• Security assumptions in DeFi are being challenged: On-chain contracts hold about $200 billion. After Drift on Solana lost $285 million, STRIDE’s security program was launched. AI vulnerability detection tools are becoming more powerful, and protocols that haven’t done formal verification will be scrutinized by depositors with more aggressive risk pricing.
• Funding is moving toward “Responsible AI”: Post-raising at a valuation implied around $380 billion, indicating investor preference for safety-oriented approaches. Whether RSP 3.0 can truly prevent misuse remains to be seen.

Benchmark scores are not everything: deployment depends on security thresholds and controllability

Enterprise procurement focuses on “can it be safely deployed,” not leaderboard rankings. A 5–10% dip in cybersecurity stocks after leaks, along with no public rebuttals from peers, shows the market values “defense/risk control” more than “capability.”

Consensus is converging: rather than asking “what else can AI do,” it’s more about “how to make it operate within boundaries.”

Key Information

• Mythos is powerful but only available for enterprise pilots: High benchmark scores combined with real zero-day samples and early enterprise access create a “high capability + strict access control” combo.
• DeFi risks are being re-priced: High-value contracts face accelerated AI attack and defense; formal verification and AI security audits will shift from bonus features to entry barriers.
• Capital and narratives are linked: Capital, talent, and resources are flowing toward “controllable strong models,” making open ecosystems and independent research increasingly marginal.
• Deployment decision logic is changing: Security thresholds, isolation capabilities, and auditability are replacing “leaderboard scores” as primary considerations in enterprise procurement.

Importance: High
Category: Model Release | AI Security | Market Impact

Conclusion: Enterprises and capital are already paying for Anthropic’s defensive positioning; builders ignoring security integration and researchers without enterprise channels face rising marginalization risks.

Judgment: From a narrative perspective, we are still in an “early but accelerating” stage; the biggest beneficiaries are builders and institutional funds with security capabilities and compliance channels. Secondary traders should closely observe shifts in security deployment and access patterns. For participants solely exposed to open ecosystems, this environment is not friendly.