In an unusually clear signal of where frontier AI is heading, OpenAI and Anthropic each unveiled highly cyber-capable models on the same day.
On September 1, OpenAI said its forthcoming model Astra is the first it has designated at the “Critical” cybersecurity threshold under its Preparedness Framework, meaning, in its words, that with the right tools it can “find previously unknown security flaws and develop ways to exploit them” without a person guiding each step. Hours earlier, Anthropic released Claude Fable 5.1 and Mythos 5.1, the latest in a Mythos line it has described as having the strongest cybersecurity capabilities of any model it has released.
For the cryptocurrency industry, whose smart contracts, bridges and wallets are critical pieces of software and infrastructure that can control or facilitate transfers of digital assets, the convergence matters, and it is worth understanding precisely rather than fearing vaguely.
Two Labs, One Capability, The Same Day
The two announcements describe similar capabilities through different in-house frameworks, so their labels are not directly comparable; nevertheless, both releases point to continued advances in AI-assisted cybersecurity.
OpenAI said Astra can, given the right access, identify and develop functional zero-day exploits in hardened systems without human intervention, and reported that in testing it discovered and used two zero-day vulnerabilities as part of an exploit chain and built working browser and operating-system compromises.
Anthropic, for its part, said Mythos 5.1 “demonstrates the strongest cyber capabilities of any model we’ve released” while, by its own assessment, still falling within the lower risk category of its Frontier Compliance Framework; Anthropic’s own description of the model focuses on advanced vulnerability discovery and cyber-defense capabilities.
Both companies also converged on the same containment strategy: gate the most dangerous capability to vetted defenders, and ship a safeguarded public version. OpenAI said Astra’s advanced cyber features will initially reach only alpha testers before expanding through a defense-focused program it calls Daybreak Blue, and that the model refuses 91.5% of cyber “jailbreak” attempts. Anthropic restricts Mythos 5.1 to vetted cyber-defenders and life-sciences organizations through trusted-access programs, while its generally available twin, Fable 5.1, applies production safeguards, an approach its predecessor implemented by routing flagged cyber requests to the weaker Claude Opus 4.8.
The financial angle is particularly relevant to crypto. Anthropic said the uplift from Mythos-level capabilities is valuable to adversaries, specifically those who could “financially gain from cyberattacks,” a category that includes attackers targeting financial and digital-asset systems.
This Collision Has Been Building For Months
The September 1 releases are the latest step in a story The Crypto Times has been tracking since spring. In April, Anthropic described withholding its more capable Mythos-class system rather than immediately releasing it publicly, a safety-driven decision that The Crypto Times previously reported. When Anthropic did ship the public version, Claude Fable 5, on June 9, The Crypto Times reported that Anthropic’s restricted testing had involved more than 10,000 vulnerabilities, underscoring the potential for the technology to function as both an offensive tool and a security-auditing system.
The capability has since reached even the cryptographic primitives that secure the internet. In late July, Anthropic’s Claude Mythos Preview had found new weaknesses in two major cryptographic algorithms: an improved attack on HAWK, a post-quantum signature scheme under U.S. review, and a faster attack on a reduced-round version of AES, the world’s most widely used encryption standard.
Anthropic stressed that neither finding affects systems in use today — HAWK is not deployed, and the AES result targets only a weakened variant — but the episode showed how AI systems are increasingly being applied to problems that previously required highly specialized security research.
As The Crypto Times framed it in its July analysis of Fable 5, the danger for crypto “is not that AI invents new hacks overnight; it is that attackers and defenders may now move at machine speed.”
Why Crypto Faces a Distinctive AI Security Risk
The reason a general cyber capability is particularly relevant to crypto comes down to how blockchain applications are built and how vulnerabilities can affect assets.
Smart contracts are self-executing programs that can custody or control digital assets. While some contracts are upgradeable, others have limited or no ability to change their core logic after deployment, meaning vulnerabilities can be difficult to remediate once users and assets depend on the code.
Cross-chain bridges, which can lock assets on one chain while enabling corresponding assets to be represented on another, have been among the significant targets in crypto security incidents. Wallets guard the credentials and keys used to authorize transactions.
In each case, an exploit can potentially translate a software vulnerability into a direct financial loss, sometimes without the intermediary steps required to monetize stolen assets in traditional financial systems.
That is not a hypothetical. In June, the privacy network Zcash was forced into an emergency network fork after the discovery of a critical counterfeiting vulnerability in its Orchard shielded pool, a flaw that, in principle, could have allowed an attacker to mint ZEC out of thin air.
The incident demonstrates why vulnerabilities in cryptographic and blockchain infrastructure can have unusually direct financial consequences. In EVMbench, a benchmark OpenAI built to evaluate AI agents on smart-contract security, the authors warn that “as AI models get better at finding bugs in software, a growing portion of the assets on crypto rails could be at risk,” and that “sufficiently capable AI systems could acquire substantial financial resources by exploiting smart contracts”, an explicit parallel to how state-linked actors have self-financed through crypto theft.
The combination of publicly inspectable code, automated execution and directly transferable assets makes crypto an important test case for the impact of increasingly capable cyber AI.
An Industry Already Under Siege
These models arrive against a backdrop of substantial crypto security losses. Chainalysis data cited put 2025 losses at roughly $6.5 billion, and on-chain security firm Blockaid reported that the first half of 2026 was the most-hacked half-year on record, with more than $1.1 billion lost across 212 incidents.
Bridges remain a leading vector: The Crypto Times reported that cross-chain bridge hacks had topped $328 million in 2026 by mid-May, led by a $292 million Kelp DAO exploit, with 2026 losses since passing $1.26 billion.
Notably, Blockaid’s report already flagged the emergence of AI-driven techniques such as prompt-injection attacks and the targeting of AI trading agents, suggesting that AI-related attack techniques are already becoming part of the crypto threat landscape.
The Other Side: AI As The Defender
The same capability is also potentially valuable to the crypto-security industry, which is why this is genuinely two-sided.
The Zcash case is instructive here too: the Ironwood remediation plan that followed the Orchard vulnerability explicitly proposes to strengthen the network’s codebase through “AI-assisted security auditing and formal verification techniques,” an example of a protocol exploring AI as part of its security response.
More broadly, AI already audits contracts and security-sensitive code, potentially allowing researchers to examine large codebases faster than conventional manual review alone. Research systems such as GPTScan reportedly found nine previously unknown vulnerabilities that human auditors missed, according to the researchers’ evaluation.
Both labs are steering their cyber capabilities toward defense first: OpenAI routes advanced access through its defense-focused Daybreak Blue program, while Anthropic has made Fable 5.1’s safeguards more precise and reserves Mythos 5.1 for vetted cyber-defenders.
For smaller DeFi teams that may have fewer resources for extensive manual security reviews, such systems could eventually become an important defensive tool. But the likely result for crypto is a faster contest between AI-assisted attackers and AI-assisted defenders.
The security challenge is asymmetric in one important respect: attackers may need to find a single exploitable weakness, while protocol developers must protect the systems they control against a much wider range of potential failures.
The Caveats
Several things should temper both the hype and the alarm. Astra is not yet generally available and its most capable cyber features are access-restricted; Anthropic’s most capable cyber model, Mythos 5.1, is likewise limited to vetted organizations, and both companies describe extensive refusal training, monitoring and red-teaming built specifically to limit misuse.
The two labs’ capability labels come from different internal frameworks and are not directly comparable, and each figure cited here is primarily based on the respective company’s own assessment or testing, pending independent scrutiny.
The cryptographic-weakness findings disclosed so far do not affect deployed systems. And crucially, neither OpenAI nor Anthropic has disclosed any instance of these models being used against a crypto system; this analysis therefore describes a capability and its potential implications, not a demonstrated AI-driven attack on a live blockchain system.
The honest summary is that the ceiling on automated exploitation is rising for attackers and defenders alike, and crypto, because vulnerabilities can be connected directly to transferable digital assets, is one of the clearest places that shift could be felt.
What To Watch
Three signals will show which way the balance tips.
First, access: how quickly the labs widen defensive programs like Daybreak Blue and Anthropic’s cyber verification track to security teams, versus how well the safeguards on the public models hold against determined jailbreaking.
Second, adoption: whether protocol teams, auditors and bridge operators build these models into continuous review the way Zcash is proposing, rather than treating audits as a one-time pre-launch checkbox.
Third, evidence: whether the first confirmed AI-discovered exploit of a live crypto system appears — and, if it does, whether it comes from a gated frontier model or from more accessible AI systems already available to attackers.
The Crypto Times will continue to track those signals, makes no market or price forecast, and takes no position on either company’s release decisions.
