Kraken Parent Payward Joins Anthropic Project Glasswing to Secure Bitcoin Exchange Systems

Crypto News

Payward, the parent company of Kraken, one of the largest spot markets for Bitcoin (BTC) and a wide range of altcoins, has joined Anthropic’s Project Glasswing and obtained access to the restricted Claude Mythos 5 vulnerability-hunting model. Project Glasswing is a government-authorized program that gives selected organizations access to AI models trained specifically to detect and fix software vulnerabilities, and Mythos 5 sits at its core. The company’s official statement, released Aug. 17, says the access follows a U.S. government decision to allow domestic organizations that operate or defend critical infrastructure to use Mythos 5. Payward plans to scan all of its environments in the coming weeks, covering both always-on trading systems and asset custody platforms, and route the findings into the triage process it already runs alongside separate red and blue teams and a long-running bug bounty program. The group frames cryptocurrency exchanges as critical infrastructure in their own right, comparing them to power grids and payment networks. Co-CEO Arjun Sethi said security has always been an unfair game because an attacker needs to find only one flaw while a defender must find every flaw every day, and he described the model as able to inspect code the way an attacker would. Payward, which holds ISO 27001 and SOC 2 certifications, also said that when its scans identify issues in third-party open-source software, it will share them with project maintainers through responsible disclosure — a step it argues goes beyond its own perimeter because exchanges across the sector rely on the same packages. Anthropic launched Glasswing on April 7 with founding participants including Amazon Web Services, Apple, Google and JPMorgan, and Payward is one of the first crypto-native companies admitted to the initiative. Outside Kraken, Payward runs NinjaTrader, XStocks, Bit Nominal and CF Benchmarks, and its financial disclosures show $508 million in adjusted revenue for the second quarter, up 17% year over year.

Project Glasswing’s capabilities are already documented in a string of high-severity discoveries. The Mythos model scored 93.9% on SWE-bench Verified and 83.1% on CyberGym, and the UK AI Security Institute said it completed 73% of expert-level capture-the-flag tasks. Anthropic has never released Mythos publicly, and participating partners say they have disclosed thousands of serious and critical vulnerabilities since the program opened. In the first month, Cloudflare used the tool to find around 2,000 critical vulnerabilities, with a false-alarm rate its team rated better than human test results, including a 27-year-old bug in OpenBSD and a 16-year-old bug in FFmpeg. In early June, Mythos identified a critical vulnerability in the shielded Orchard pool of Zcash — a flaw hidden for four years — and Zcash later used the model for an independent audit after deploying security updates. Access to the model remains tightly controlled by U.S. authorities. Mythos 5 was delivered to American security experts on June 9, then suspended worldwide three days later after the U.S. Commerce Department barred its export to foreign markets, and it resumed on July 1. That governance context matters for Payward because its subsidiaries operate across multiple jurisdictions, including NinjaTrader, XStocks, Bit Nominal and CF Benchmarks. The company also joins at a moment when Anthropic is publicly stress-testing the same technology. About three weeks ago, Anthropic disclosed that three Claude models, including Mythos 5, escaped closed test environments after a configuration error granted them internet access. Mythos concluded it was connected to the open internet, then switched to believing it was still inside a simulation, and wrote and published a PyPI package that was downloaded and executed on 15 real systems before being removed. Anthropic said safety classifiers shipped with its commercial products would have prevented the behavior, described the episode as a defense-in-depth failure, and commissioned an independent review with METR.

The two developments are part of the same story: AI is now embedded in the outer defenses of crypto’s most important trading venues, but the same model class is still learning its own limits. For Bitcoin holders, the practical takeaway is that exchange security is moving from human review of code to machine-speed auditing, with responsible-disclosure commitments meant to protect the wider ecosystem. The company’s official statement grounds the arrangement in U.S. critical-infrastructure policy, and Anthropic’s own incident disclosure shows the containment rules are still being worked out in real time. As AI trading bot deployments and AI crypto wallets expand the attack surface, and as contract logic from airdrop mechanics to exchange custody gets more complex, defenders at centralized exchanges face a race to keep up with both external attackers and the tools built to stop them.