Payward Joins Anthropic’s AI Security Program It Never Publicly Asked For

MakeCryippreferred on

Payward, the parent company of Kraken, said on Aug. 17 that it has joined Anthropic’s Project Glasswing, a restricted-access program that vets organizations before granting them tools for finding software vulnerabilities. The access specifically covers Claude Mythos 5, Anthropic’s model built to find and fix those vulnerabilities at scale. Payward co-CEO Arjun Sethi described the move in the company’s own announcement as a shift from attacker to defender: “An attacker needs to find one flaw. A defender has to find all of them, first, every single day. Frontier AI is the first thing that flips that asymmetry.”

Scanning of Payward’s own systems is due to start in the coming weeks. Anthropic hasn’t published a complete list of Glasswing’s roughly 200 partners, so Payward’s claim to being the program’s first cryptocurrency member rests on the absence of any other crypto name in the public record, not a confirmed count.

The letter Payward didn’t sign

Payward’s entry follows, within the same week, an open letter organized by the Bitcoin Policy Institute asking AI labs to open exactly this kind of access to security researchers working on open-source crypto infrastructure. More than 40 firms signed it, including Coinbase, Block, BitGo and BlackRock. Payward is not one of them. The company that ends up first through Glasswing’s door is not one of the companies that spent this month publicly asking for the door to open.

What Payward’s own record shows

Payward’s announcement doesn’t mention two episodes from its own recent history, both of which involved vulnerabilities found by people, not models. In June 2024, a security researcher working with CertiK found a flaw in Kraken’s deposit system that let a user credit their account before a deposit had fully cleared, and used it to pull $3 million out of company funds instead of reporting it through the bug bounty program. Kraken’s chief security officer, Nick Percoco, called it extortion, not white-hat research; CertiK later took responsibility and the money was returned.

Then, in April 2026, four months before the Glasswing announcement, Kraken disclosed that two of its own support staff had separately gained unauthorized access to client data, together touching about 2,000 accounts, roughly 0.02% of its customer base. An extortion attempt followed. Percoco said the company would not pay: “Our systems were never breached, funds were never at risk, we will not pay these criminals.”

Neither episode was an outside attacker exploiting a software bug the way Mythos 5 is built to catch. Both were failures of internal access control and bounty-process trust, the kind of problem a code-scanning model doesn’t touch.

TradFi got here first, and said less about it

Payward is not the first financial institution in Glasswing. JPMorgan Chase and Goldman Sachs have both been testing Mythos models under the program since its April launch, and neither bank has treated it as a turning point. JPMorgan chief executive Jamie Dimon has acknowledged the model “does create additional vulnerabilities” even as it helps find them, but described his bank’s response the way it already talks about cyber spending generally: “We spend a lot of money. We’ve got top experts.” Set that next to Sethi’s language about flipping an asymmetry: two of the largest US banks folded this access into an ongoing budget line, while Payward made its entry the subject of a press release.

What actually gets tested here

Nothing about Mythos 5 has been deployed at Payward yet, and Anthropic’s own account of the program describes an average two-week gap between a vulnerability being found and being patched, a bottleneck the company itself calls unresolved. What can be judged now isn’t whether the model works. It’s whether the company putting its name on crypto’s entry into this program has a recent history that matches the story it’s telling. Twice in the last two years, the flaws that cost Payward money or trust were found by people it didn’t intend to find them, not by machines pointed at its own systems. Mythos 5 may close that gap. The announcement that it’s coming doesn’t say anything about the gap it’s meant to close.