In April, Anthropic did something the AI industry had not seen in years: it built a model so capable it refused to release it. The company unveiled a “Mythos-class” tier above its Opus models, warned that the system could discover and weaponize software vulnerabilities at scale, and said it had no plans for public release due to insufficient safeguards.
Observers called it the most prominent safety-driven withholding of a frontier model since OpenAI held back GPT-2 in 2019.
On June 9, that stance flipped. Anthropic publicly released Claude Fable 5, the first Mythos-class model made generally available, arguing its new safeguards are now robust enough. It is the most capable model the company has ever shipped to the public—and for crypto, the timing and capabilities are impossible to ignore.
What made the Mythos tier too dangerous to release is exactly what makes it matter for crypto. These models excel at discovering and exploiting software vulnerabilities and at “agentic hacking.”
As mentioned in the company’s official document, the model family helped partners surface more than 10,000 vulnerabilities in restricted deployment inside Project Glasswing, Anthropic’s cyber-defense program.
In a world where smart contracts hold billions and code is money, an AI that can probe protocols at machine speed is simultaneously the best auditor a project could dream of and the most capable adversary it has ever faced.
Why crypto should care
Crypto enters this moment with fresh scars. According to TRM Labs’ 2026 Crypto Crime Report, illicit actors stole $2.87 billion across nearly 150 hacks and exploits in 2025—making it the worst year on record for crypto theft by value. The largest single event, the $1.46 billion Bybit breach in February 2025, accounted for over half of the year’s total and remains the biggest crypto heist in history.
With total scam-and-fraud losses running far higher, 2026 is already off to a painful start, with North Korean-linked groups alone stole approximately $577 million in the first four months (76% of all hack losses YTD), with additional incidents pushing the early-year total higher.
This risk sits against a much larger backdrop of roughly $120–150 billion in DeFi currently residing in smart contracts, bridges, lending protocols, and liquid staking pools. This capital is most directly exposed to machine-speed vulnerability discovery and agentic exploits.
Security researchers had already flagged frontier AI as the next major inflection point. On-chain protocol code has grown harder to exploit directly, pushing attackers toward social engineering and emerging on-chain AI agents. Analysts warned that in 2026, AI would accelerate both defense and offense at once: machine-speed monitoring for defenders and dramatically faster vulnerability research, exploit development, and scaled social engineering for attackers. A model with Fable 5’s profile is precisely what both sides have been preparing for.
The defensive upside is immediate and concrete. In early testing, Fable 5 compressed months of engineering work into days. Stripe reported it completed a codebase-wide migration across a 50-million-line repository in a single day. Applied to Solidity or Rust smart contracts, this level of throughput could transform auditing, formal verification, and the continuous automated review that most projects still skip.
Locked behind the vault
The launch is deliberately engineered around misuse risk. The fully capable, cyber-unlocked version—Claude Mythos 5—remains restricted to vetted cyber defenders and critical infrastructure providers through Project Glasswing (in collaboration with the US government). It uses the same underlying model as Fable 5, with safeguards lifted in targeted areas.
Public Fable 5, by contrast, ships with classifiers that intercept queries on cybersecurity, biology, chemistry, or model distillation and route them to the less-capable Claude Opus 4.8. Anthropic reports the guardrails block meaningful progress on offensive cyber tasks, more than 95% of sessions trigger no fallback, and a 1,000-hour external bug bounty found no universal jailbreaks.
One blocked category stands out for crypto readers concerned about capability sprawl is distillation. Anthropic specifically flagged the risk that leaking capabilities could let near-frontier power spread into unsafeguarded models. In a sector where forks and clones proliferate overnight, that concern is pointed—the guardrails hold only as long as this remains the sole model in its class.
Beyond exploitation, Fable 5’s standout trait is autonomy. It maintains coherence across millions of tokens in long-running tasks and improves its own work from notes and memory. On benchmarks it leads:
- SWE-Bench Pro (agentic software engineering): 80.3% — well ahead of Opus 4.8 (69.2%) and GPT-5.5 (58.6%).
- FrontierCode Diamond (high-quality, maintainable agentic coding): 29.3% — more than double Opus 4.8’s 13.4%.
These gains shine on long-horizon and complex tasks, exactly the profile needed for deep smart-contract analysis or autonomous agent orchestration.
The double-edged agent thesis
That autonomy directly fuels the AI-agent narrative already powering a corner of the crypto market—autonomous on-chain agents that execute strategies, manage positions, and interact with protocols without human intervention.
Yet the same January analyses that highlighted AI’s defensive potential also flagged on-chain AI agents as a brand-new attack surface: faster and more powerful than humans, but uniquely vulnerable when permissions and control paths are sloppy. More capable models raise both the ceiling on what autonomous crypto systems can achieve and the potential damage when one is compromised or manipulated.
On the markets side, Fable 5 also delivers category-leading performance on finance and trading-analysis benchmarks, including top scores on senior-level finance reasoning tests—another signal for the quant and analysis tooling increasingly integrated into digital-asset operations.
The race is on
Projects and protocols should start testing Fable 5 on their codebases now. The era of AI-augmented auditing has arrived, but so has the era of AI-augmented attacks. The winners in 2026 and beyond will be those who use these tools aggressively for defense while building robust safeguards around the autonomous agents they deploy.
The model is here. The question is who will use it first—and smartest.
Also read: One Laptop, $36 Million, and a Token Collapse: Inside the Humanity Protocol Exploit