This is outside our normal publishing schedule and a little different from the usual structure. On April 7, Anthropic (the company behind Claude) released a security assessment with direct implications for credit union cybersecurity and vendor risk. Let’s talk about it.

In November 2023, a ransomware attack hit Ongoing Operations, a cloud services provider owned by Trellance. Within hours, 60 credit unions went dark. Members couldn't access accounts. Staff couldn't process transactions. The attack targeted a vendor, and every credit union on that vendor went down with it.

In August 2025, the same pattern repeated at Marquis Software Solutions. A flaw in a SonicWall firewall exposed the data of more than 672,000 members across 74 banks and credit unions.

On April 7, 2026, Anthropic published a security assessment of their newest AI model.

Anthropic’s new AI model, Claude Mythos Preview, is the first model in its new Capybara tier above Opus, and under controlled lab conditions it was able to identify and exploit previously unknown security flaws in major operating systems and web browsers.

In a standardized test, the previous best AI model produced two working attacks. Mythos produced 181. It found bugs hiding in widely used software for 16, 17, and 27 years, flaws that every human security expert and every automated scanning tool had missed.

The report also describes flaws in the encryption software that protects data moving between your members and their banking apps (bugs in the code, not breaks in the encryption math itself).

Anthropic employees with no formal security training asked Mythos to find vulnerabilities overnight, in Anthropic's own lab environment, and woke up to complete working attacks. Individual successful runs cost as little as $50 in compute, though finding a given vulnerability may require thousands of runs at far higher total cost. The expertise and cost barriers to finding these bugs are dropping fast.

Some researchers urge caution about the scope of these claims. Dr. Heidy Khlaaf, chief AI scientist at the AI Now Institute, has noted the absence of false positive rates and independent verification in the report.

Lab benchmarks are not the same as reliable, scalable attack workflows in the wild. The gap between demonstrated capability and operational threat is real. It is narrowing. Anthropic's decision to restrict access rather than release publicly is a meaningful step, but it does not change the trajectory.

The National Credit Union Administration's (NCUA) 2025 Cybersecurity and Credit Union System Resilience Report to Congress found that "approximately 73 percent of all reported incidents were related to the use or involvement of a third party." The SonicWall-style flaws that enabled the Trellance and Marquis breaches are the same class of bugs Mythos uncovers at scale.

Much of the software credit union vendors run on was written decades ago in programming languages that are especially prone to the kinds of bugs Mythos finds best (55% of banks cite legacy systems as their top barrier to transformation…I don't have an equivalent credit union figure, but the pattern is familiar).

Those systems are tightly coupled to vendor patch cycles that run quarterly or annually. The Cybersecurity and Infrastructure Security Agency (CISA) requires federal agencies to remediate critical vulnerabilities on internet-facing systems within 15 calendar days of detection, a standard widely adopted as an industry benchmark. The Mythos report documents attacks that would take human security experts weeks to build. The model did them in hours.

As of April 9, Anthropic reports that fewer than 1% of the potential vulnerabilities it has discovered so far have been fully patched by maintainers.

Anthropic isn't releasing Mythos publicly. They've restricted it to Project Glasswing, a defensive coalition of named launch partners including JPMorgan Chase, Apple, Google, Microsoft, and CrowdStrike, plus over 40 additional organizations. Logan Graham, Anthropic's frontier red team lead, warned in Wired: "We need to prepare now for a world where these capabilities are broadly available in 6, 12, 24 months."

Wharton professor Ethan Mollick put it more bluntly: "Based on historical trends in AI they have about six to nine months until those capabilities become widely diffused to bad actors."

No credit union, credit union service organization (CUSO), or community bank is among the named Glasswing partners. JPMorgan has thousands of security staff. A $500 million credit union has one or two IT people. The NCUA has been requesting vendor examination authority from Congress for years without success. Credit unions are entering this window dependent on vendors whose security practices their own regulator cannot examine.

For example:

Schedule a board briefing on AI-accelerated cyber threats. The NCUA already lists cybersecurity as a governance priority. The Mythos report changes what "adequate" looks like.

Review vendor contracts for security assurance clauses, patching commitments, and incident notification requirements. After Marquis, the question of who bears liability when a vendor breach exposes your members' data belongs in your contracts.

Monitor the vulnerability disclosures coming out of Project Glasswing over the next 90 days. Anthropic committed to notifying affected software makers and giving them 90 days to fix the bugs before going public, starting April 7. Many of those bugs will be in software your vendors depend on.

Security defenses built on friction (tedious but not technically hard barriers) are the first to fall against AI-assisted attackers. Anthropic's report says so explicitly. What was labor-intensive enough to deter a human attacker is automated away. The defenses that hold are the ones with actual technical barriers behind them.

The Trellance attack took down 60 credit unions through one vendor. The tools to find the next vulnerability like it are getting faster, cheaper, and more accessible every month. The window to get ahead of that curve is open now. How long it stays open depends on how fast these capabilities spread.

The Anthropic red team report is available in full. Project Glasswing details are at anthropic.com/glasswing.