Payward, Kraken’s parent company, has joined Anthropic’s restricted Project Glasswing and will use Anthropic’s Claude Mythos Preview for defensive cybersecurity work. The goal is simple: find software weaknesses before attackers do.
- Payward joins Anthropic’s Project Glasswing
- Claude Mythos Preview will scan for vulnerabilities
- Human review still decides what gets fixed
- AI security tools bring real upside, and real misuse risk
The pitch is easy to understand. Let an advanced model read code and system behavior at machine speed, flag weak spots, and help security teams move faster. The catch is just as obvious: tools that can find bugs can also help build exploits if they end up in the wrong hands. Crypto has enough problems without handing attackers a nicer wrench.
Payward co-chief executive Arjun Sethi put it this way:
“A model can read every line of code the way an attacker would, at machine scale, so we find the flaw before anyone can build the exploit, ” Sethi said.
What Payward is doing
Payward says it has joined Anthropic’s Project Glasswing, a restricted-access cybersecurity program for vetted organizations, and gained access to Claude Mythos Preview. The company plans to use it to scan its environments for vulnerabilities in the coming weeks.
That does not mean an AI gets free rein to rewrite code or push fixes on its own. Findings will still go through Payward’s existing security review process before anything is changed. That human checkpoint matters. AI can surface a lot of possible issues, but security teams still have to separate real bugs from false positives, confirm the risk, and make sure a fix does not break something else in production.
Payward also said it will share validated third-party findings with the relevant open source project maintainers. That is coordinated disclosure in plain English: report the flaw privately first, give maintainers time to patch, and avoid turning an unfixed bug into public sport before the code is safe.
Several details were not disclosed, including the cost of access, the deployment schedule, and which systems will be scanned first. Payward also did not say whether the model will inspect production systems, isolated copies of source code, or controlled testing environments. Those are not small omissions. They determine how useful the setup is, and how much risk it introduces.
Why this matters
This is not just a Kraken footnote. It is a sign that AI-assisted security scanning is moving from hype deck territory into real infrastructure work.
Anthropic launched Project Glasswing as a restricted program for selected organizations and later expanded it to approximately 150 organizations across more than 15 countries, according to the company. Initial participants included Amazon Web Services, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, Nvidia, Palo Alto Networks, and the Linux Foundation.
Anthropic says the reason for the restriction is straightforward: its strongest cybersecurity models can help defenders, but they can also be used offensively to find vulnerabilities and produce working exploits. That is exactly why the program is not open to everyone with a keyboard and a bad idea.
The company has also pushed the broader case that AI can accelerate vulnerability discovery at scale. In Anthropic’s own materials, Claude Mythos Preview reportedly found more than 10, 000 flaws classified as high or critical severity across widely used software. Its coordinated disclosure dashboard showed 1, 596 vulnerabilities across 281 open source projects as of May 22.
Anthropic said the dashboard showed a 90.8% true positive rate among 1, 900 candidates reviewed by external security firms. In other words, most of what the model flagged was real, but not all of it was. That distinction is the whole ballgame in security: finding issues is useful; flooding teams with noise is not.
Anthropic also said only 97 listed findings had been patched upstream at the time of that dashboard update, and 88 had received a public advisory identifier. “Upstream” just means the original project maintainers, rather than a downstream company or user, applied the fix in the main codebase.
The upside, and the obvious downside
There is a real defensive payoff here. AI tools can scan more code, more configs, and more dependencies than a human team can realistically review line by line. That includes the boring but dangerous stuff: exposed secrets, unsafe input handling, misconfigured cloud permissions, stale libraries, and other mistakes that tend to sit quietly until somebody starts taking money or data.
For crypto platforms, that matters even more. Exchanges and custody systems are high-value targets that never really sleep. A missed flaw can become a very expensive lesson very quickly.
But the dual-use problem is not theoretical. The same model that helps a defender spot a weak authentication flow can help an attacker look for the same weakness. Anthropic itself has been blunt about that risk, which is why access is restricted and not generally available.
Human review is still mandatory for another reason: AI security tools can be wrong in ways that are annoyingly expensive. They can miss context, misread code paths, or flag issues that look serious but are not exploitable. A machine can be very confident and still be very wrong. That is not a bug unique to AI; it is just the oldest problem in security wearing a shinier hat.
Anthropic also requires Mythos 5 customers to accept thirty day data retention for safety monitoring. That may be a reasonable guardrail, but it raises fair questions about what data gets submitted, how sensitive information is separated, and whether customer information stays out of the scanning process altogether. Payward has not explained that part.
How Anthropic frames the bigger picture
Anthropic’s Glasswing materials make a broader argument: finding vulnerabilities is no longer the hardest part. Verifying them, coordinating disclosure, and actually getting patches merged is where things bog down.
That is the unglamorous truth most AI-security marketing skips over. It is nice to say a model found thousands of flaws. It is another thing entirely to prove they are real, notify the right maintainers, wait for fixes, and then roll those fixes out without creating new problems.
Anthropic also says many of the organizations in the program work in critical infrastructure sectors such as power, water, healthcare, communications, and hardware, and that for most partners, a major attack could affect more than 100 million people. That is Anthropic’s framing, not a neutral law of physics, but the point is clear: the blast radius of a missed bug can be huge.
Payward itself has argued that digital asset platforms should be treated as critical infrastructure. That may be a fair description in practical terms, these systems do handle valuable assets and continuous transfers, but it is still Payward’s position, not an official designation. Calling yourself critical infrastructure does not make the regulators appear out of thin air.
Why crypto security teams are paying attention
Kraken has long tried to position itself as one of the more serious players in the exchange business, and this move fits that image. The company is treating security as an ongoing engineering problem, not a box to tick once and forget until the next incident.
That is the right instinct. Crypto infrastructure is a magnet for attackers because the prize is obvious and the targets are permanent. If AI can help find weak points faster, that is a real advantage. It is not a magic shield, but it is better than pretending manual review alone can keep up forever.
There is also a supply-chain angle worth keeping in view. Payward said it will share validated findings with open source maintainers, which matters because one flaw in a shared library can ripple far beyond a single company. In crypto, one bad dependency can become everybody’s problem in a hurry.
And yes, Kraken has dealt with security drama before. The exchange previously patched a deposit flaw that researchers used to withdraw nearly $3 million, then later recovered the funds after a public disagreement with CertiK. That episode is a reminder that these issues are not abstract. In crypto, a bug is not just a bug. It is often a direct line to money.
That is why the details here matter so much. A flashy AI announcement is not proof of better security. What matters is whether the tool actually helps teams find real flaws, validate them, patch them cleanly, and do it without exposing sensitive data or creating a new attack surface.
Key questions and takeaways
-
What did Payward join?
Payward joined Anthropic’s Project Glasswing, a restricted cybersecurity program for vetted organizations. -
What model is Payward using?
Payward said it has access to Claude Mythos Preview for defensive security scanning. -
What will the model be used for?
It will scan Payward’s environments for software vulnerabilities, with human review still required before any fixes are made. -
Why is access restricted?
Anthropic says these models can help defenders, but they can also help attackers find flaws and build exploits. -
Will the AI fix bugs automatically?
No. Payward says findings go through its normal security review process first. -
Why does coordinated disclosure matter?
It gives open source maintainers time to patch vulnerabilities before details are made public. -
What is the main risk here?
The same dual-use tooling that strengthens defense can also be abused if access is leaked or misused. -
What is the bigger crypto takeaway?
AI is becoming a real security tool for crypto infrastructure, but it only helps if humans stay in control and the disclosure process stays disciplined.
Payward’s move is sensible. It is also a glimpse of where security is headed: more AI-assisted review, more automation in vulnerability discovery, and more pressure on teams to actually ship fixes instead of admiring their dashboards. Useful progress, sure. But in crypto, every powerful defense tool comes with a matching offense manual if it leaks, gets abused, or lands in the wrong hands.
Further reading
A few extra pieces on AI security, Glasswing, and Payward’s broader move into the defensive trenches.
- Payward taps Claude Mythos 5 for crypto security
- Cybersecurity in the Age of AI
- Error extracting content
- Claude Mythos #2: Cybersecurity and Project Glasswing
- Claude Mythos
- Kraken parent Payward joins Anthropic's Project Glasswing
- Kraken Parent Payward Revenue Rises 3% to $507M as
- US Military Runs Bitcoin Node for Cybersecurity Tests, Not
- U.S. Military Tests Bitcoin Node for Cybersecurity, Not
