On April 30, 2026, Anthropic launched Claude Security into public beta. At first glance, it seems like a routine release of another corporate tool. In reality, this is the first mass attempt to shift the work of information security specialists onto a language model: scanning repositories, identifying vulnerabilities with data chain indications, assessing severity with confidence levels, providing ready-made patches, and exporting findings to Slack and Jira. Prior to this, the product had been in a closed preview for six months under the name Claude Code Security and, according to Anthropic, identified vulnerabilities that traditional scanners had missed for years. Now it is available to all subscribers of Claude Enterprise.
At West Star Ltd, we work with business process automation and integrations for Kazakhstan's B2B sector, and we see how this shift will manifest in the local reality. And this reality is more alarming than it appears from the polished releases — because AI tools for attacks are spreading faster than those for defense.
What Has Changed
Claude Security is built on the Opus 4.7 model and does not operate like a typical static analyzer. Instead of searching for signatures, it reasons over the entire codebase as a human researcher would: tracing data flows, reading the source code of dependencies, analyzing component interactions, and synthesizing this into a comprehensive picture. Each identified vulnerability comes with a confidence assessment, reproduction steps, and a suggested patch. Through the integrated Claude Code, the session transforms from "we found a problem — opened a ticket — waiting a week" to "found — fixed — verified" in one go.
The product's partner network showcases Anthropic's ambitions. CrowdStrike, Palo Alto Networks, SentinelOne, Trend Micro, and Wiz are embedding Opus 4.7 into their platforms. Accenture, BCG, Deloitte, Infosys, and PwC are helping to deploy solutions on the client side. This is not a niche product for enthusiasts — it is an attempt to become the new standard in corporate AppSec.
Simultaneously, Anthropic rolled out Project Glasswing with the Claude Mythos Preview model — a closed version that the company positions as capable of finding and exploiting vulnerabilities at or above the level of elite human experts. According to their own data from the Mythos Preview review, independent specialists agreed with the model's severity assessment 89% of the time accurately and 98% of the time with a deviation of no more than one level. If extrapolated to the rest of the findings, this refers to thousands of critical vulnerabilities that the model discovered in real production code, including operating system kernels.
Here, the important aspect is not the number, but the vector. Anthropic explicitly states in the announcement: AI compresses the time between vulnerability discovery and exploitation, and the only adequate response is to provide defenders with tools of the same class that attackers will have. The window in which "we have a hole in our code, but no one has found it" is closing.
Why This Matters More for Kazakhstan Than It Seems
Local statistics look like this. According to the State Technical Service, in 2024 Kazakhstan recorded over 41,000 information security incidents — a 2.6-fold increase compared to 2023. In 2025, a database containing personal data of 16.3 million Kazakhs — essentially the entire adult population of the country: IINs, addresses, phone numbers, birth dates, medical data — was leaked online. According to TSARKA, this is a collection of previously stolen data consolidated by IIN. The source is private systems; no attacks on government resources were recorded in this incident.
In the Global Cybersecurity Index by the ITU, Kazakhstan ranks 78th — third in its group of 29 countries. The PwC 2025 Global Digital Trust Insights report, which included Kazakh respondents, records two things simultaneously: only 2% of companies worldwide have built a corporate-wide cybersecurity resilience system, while 67% of security leaders acknowledge that generative AI has expanded the attack surface. This means the threat is recognized, but defensive measures are lagging.
Regulators in the country are already taking action. The new "On Banks" law of 2026 requires biometric identification, anti-fraud systems, and continuous monitoring. Starting in 2026, biometric data will be required for SIM card registration. The Anti-Fraud Center prevented about 100,000 incidents in 2025, saving banks' clients around 3 billion tenge. This is the right direction — but it only addresses one market segment (banks and telecom operators) and hardly touches the thousands of Kazakhstani small and medium-sized businesses that operate custom or legacy code that has never undergone a security audit.
It is precisely these companies that will bear the brunt. When the time from the discovery of a CVE in a popular library to mass exploitation shrinks from weeks to hours, the presence or absence of systematic scans ceases to be a matter of maturity — it becomes a matter of survival.
What Changes for Developers and Business Owners
The first change is the economics of vulnerability discovery. A pentest audit by an external team for a typical SaaS product in Kazakhstan costs from 2–3 million tenge for a one-time project, and companies usually conduct it once a year — if they do at all. Claude Security, through a subscription to Claude Enterprise (starting at $150 per seat per month + usage based on API rates), allows scheduled scans across all code, all branches, and on every significant change. This is not a replacement for pentesting — the model does not simulate social engineering, does not work with infrastructure, and does not attempt WAF bypasses. But it removes a significant portion of the work related to auditing the code itself.
The second change is who can afford SecOps. Previously, building an internal practice required an AppSec engineer with a salary of 1.5–2 million tenge per month in the Kazakh market, plus 6–12 months to establish processes. Now, a team of two developers can connect the tool in an evening, receive the first report the next day, and start addressing findings iteratively. This does not make them a full-fledged SOC, but it fundamentally changes the entry point.
The third change is the pressure on those who sold "findings as a product." Analyst Mitch Ashley from The Futurum Group commented for DevOps.com: vendors whose value lived in the gap between discovery and fix are losing that gap. When the model immediately offers a working patch, you pay not for a list of problems, but for their resolution.
For Kazakhstani businesses, this means that the window for building their own security practice is open right now. In two to three years, when Mythos-class models become widely available and attackers begin to exploit them in the mass segment, playing catch-up will be an order of magnitude more expensive.
Where This Picture Has Weak Spots
Ignoring the limitations of this new class of tools means preparing for an unpleasant surprise. They exist, and they are significant.
Dependence on the model and vendor. Claude Security is only available as part of the Enterprise subscription and works only with Opus 4.7 from Anthropic. Local deployment is not possible. For companies working with critical infrastructure or personal data of Kazakh citizens, this raises questions about cross-border data transfer and compliance with regulatory requirements. Cloud storage is cited as the number one threat in PwC reports (42% of executives), and the solution of "let's hand over all our code to an American vendor's cloud for analysis" is a compromise in any case.
False positives and "AI fatigue" effect. Anthropic claims that multi-step validation reduces noise, but initial user reports show that on the first run, the model finds dozens and hundreds of warnings, and the team drowns in triage. Without a well-established filtering process and a clear "owner of fixes," the tool quickly turns into noise that is ignored.
The invisible side of vulnerabilities. Most leaks in Kazakhstan and worldwide occur not due to a hole in the code, but due to misconfigured access, a compromised administrator, incorrect S3 bucket configuration, or human error. According to global statistics, 46% of breaches are internal threats. Claude Security works with code, not with people or production environment configurations. If a company has an admin account lying in a public chat, no repository scanner will help.
Arms race on the attacker's side. Anthropic honestly admits: the same class of models that defenders are receiving now is already available to attackers — albeit in the form of less advanced analogs or leaked weights. The time from CVE publication to mass exploitation in 2026 is already measured in hours, not days. This means that having a scanner does not guarantee protection — the ability to quickly roll patches into production is needed. For a typical Kazakhstani company, where a release every two weeks is the norm, such a pace is currently unattainable.
A new class of risks — IP leaks through the tool itself. When the model reads all your code, including business logic and integrations, any compromise of the subscription or error in access policies turns into a leak of the entire codebase. Anthropic claims it does not use the data of Enterprise clients for training, but policy is not a technical guarantee.
Practical Conclusion
The shift that occurred in April-May 2026 is not about "a new product from Anthropic." It is about AppSec ceasing to be a separate road profession and becoming a utility — like a linter or CI/CD pipeline today. In a year, the absence of automated code scans in a company will be perceived as the absence of backups: formally, it is possible to operate, but at the first serious problem, there will be no explanations to clients or regulators.
What makes sense for Kazakhstani businesses to do right now, without waiting for the next high-profile incident:
- Conduct an inventory of code and dependencies. Understand where your source code is located, what external dependencies it has, and who has access to it. Without this, any scanner will work blindly.
- Launch a pilot on one project — not on everything. Choose the most critical service (where money or personal data of clients are involved), run it through Claude Security or an equivalent, refine the triage and fix process. Only then scale up.
- Close basic hygiene gaps before spending money on advanced tools. Multi-factor authentication, access segmentation, regular key and token rotation. According to GTS, the main incidents in Kazakhstan are basic level attacks, not APTs.
- Budget for 2026 not for a one-time pentest, but for a subscription to continuous scanning tools. The cost per month is comparable to the salary of one junior developer, while the coverage is greater than that of an external auditor once a year.
In our products — AI accountant, OData Hub, warehouse integrations with 1C — the issue of code and data security arises every time client integrations are discussed. The approach that works: do not try to build perfect protection from day one, but establish a process where every significant change undergoes automated checks. This used to be expensive. Now, in May 2026, it has become a matter of discipline, not budget. And this, perhaps, is the main change.