Why Wall Street panicked over AI and why it was an overreaction
U.S.-based AI giant Anthropic canceled the public release of its new Claude Mythos Preview model, citing its unprecedented capabilities in software vulnerability detection. On April 8, the U.S. Department of the Treasury and the Federal Reserve convened an emergency meeting with CEOs of major banks. Cybersecurity stocks fell by 7% to 9%, even as other factors drove the market lower. According to analysts, the cybersecurity sector reacted disproportionately to Anthropic’s news. The media rushed to declare the «end of the zero-day era.»
Of course, this is a serious signal and the financial sector should heed it. However, it is not what the media is claiming. The model only works if the source code is available, but much cheaper tools can also do this. The real signal is different: regulators have publicly admitted, for the first time, that AI not only reduces the cost of attacks but also significantly simplifies them and potentially enables automated, multi-stage attacks without human intervention. Although the situation is manageable, the window for a calm and systemic response is limited.
What really happened
Mythos is indeed capable of identifying software vulnerabilities better than previous models. The specific results are impressive: a 27-year-old bug in OpenBSD, a 16-year-old one in FFmpeg and a vulnerability in FreeBSD that automated tools had unsuccessfully searched for 5 million times. These are real findings in real code.
However, a critical context is missing from most publications.
All these results were obtained using static analysis, meaning the models were provided with the program’s complete source code and asked to find errors. This is like giving an auditor full access to the entire accounting system and asking if there are any violations. It is still valuable, but fundamentally different from attacking a system from the outside. In dynamic testing, where source code is unavailable, Anthropic itself reports significantly more modest results.
The Firefox tests, which were particularly widely cited by the media, were conducted without a sandbox enabled, meaning they were theoretical vulnerabilities under controlled conditions, not a ready-made attack tool for a real user.
Three facts that weren’t included in the press release
It’s important to be honest about Anthropic’s argument. The company didn’t claim that only Mythos can find vulnerabilities. They were talking about something else. The model is capable of performing the task autonomously and continuously at a scale beyond the abilities of human or less powerful AI systems. It’s about the attack’s scalability, not the detection of a specific bug. This is the core of their warning.
The first fact. The independent research company AISLE took the same vulnerabilities from Anthropic’s press release and reproduced them using small, cheap and open-source models, including a model costing $0.11 per million tokens. All eight were successful when given isolated code with context. It’s worth noting that AISLE is a startup in the AI security space, competing in the same niche. Their findings are noteworthy, but the source is not neutral.
AISLE demonstrated that even cheap models can find a specific bug in isolated code with context. But this is not what Anthropic is warning about. The real advantage lies in its scale and autonomy: Mythos can independently traverse an entire codebase, build a chain of several vulnerabilities and generate a ready-made exploit — a tool that uses the identified vulnerability to carry out a real attack, without human intervention. This is a task of fundamentally different complexity. There is no data yet on whether small models are capable of this.
The second fact. As Epoch AI analyst Ramez Naam showed, after proper tuning of Mythos metrics, it is only slightly above GPT-5.4 in performance. So, we’re not talking about a discrete leap; it’s simply an evolution within the context of overall industry progress.
The third fact. Project Glasswing, an initiative under which Anthropic loaned $100 million to 12 of the largest U.S. corporations to use its model, is both a defensive initiative and, according to some analysts, an attempt to establish a dominant position in the corporate cybersecurity sector at a time when demand for such solutions is at its peak. There’s no direct evidence of this, but the timing coinciding with the Pentagon controversy is noteworthy. When Apple, Microsoft and Google publicly discuss the threat, it’s important to consider that they gained access to the tool on favorable terms and have a clear interest in being part of the story.
So why did Wall Street panic?
The reaction of financial regulators is illustrative, but the reason behind it is different from the one everyone is trying to exploit for hype.
Scott Bessent and Jerome Powell brought together the bank CEOs not because Mythos is something invincible. Their decision followed a leading AI lab’s public announcement that it has created a tool that automates what previously required a team of highly qualified specialists and several months of work. This is a precedent. Of course, financial regulators are right to respond to this precedent rather than wait for proof.
The panic reflects an institutional fear of uncertainty, not a proven scale of the threat. It’s a normal reaction to a real signal, which is a signal about where the industry is going, not a specific catastrophe.
What does this mean for Kazakhstan’s fintech, the IT industry and beyond?
I agree with most of the recommendations circulating in the industry. But it is important to understand why they’re true, rather than accept them simply because of hype with no proof.
Whether it is Mythos or something else, the cost of automated vulnerability scanning has been declining every year. This is a steady trend that was in place before April 2026 and will continue. Any specific model is just a point on this curve.
This trend leads to specific conclusions. Code that hasn’t been audited in the last few years needs to be audited, not because Mythos will hack it tomorrow, but because auditing tools have become more accessible to everyone, including attackers. Dependence on dozens of SaaS providers creates an attack surface that is difficult to control. On-premise language models for auditing your own code are a reasonable investment, justifiable without apocalyptic rhetoric.
Organizations outside Project Glasswing — which includes our entire region, independent developers and most European companies — will not have access to Mythos’s security capabilities in the coming months. This is a real asymmetry. However, the answer is to develop your own expertise rather than wait for invitations from American consortia.
What to do right now
Audit your codebase. Check components older than five years, especially those handling network traffic, authentication and transactions. Open-source tools like Semgrep and CodeQL are already available and quite mature. If you haven’t had an audit in the last two years, do it now.
Audit SaaS dependencies. Create an inventory of all third-party services with access to your infrastructure. For critical integrations, minimize privileges to the bare minimum.
On-premise LLM for internal auditing. Deploy a local language model, such as Llama, Mistral, or a similar model, to regularly analyze your own code. This will automate routine searches for common vulnerabilities without transferring your codebase to third-party services.
Participate in open-source security. If your products depend on open-source components — and they do — find resources to audit them. A single engineer systematically reporting vulnerabilities in the projects you use is sufficient.
Update your threat model. Accept the fact that automated vulnerability scanning of your code has become significantly cheaper. Your security priorities should reflect this.
Panic generates isolated patches. A systemic response to reducing the cost of attacks is a program, not a to-do list. When Anthropic’s first public report is released in July 2026, that will be a good time to launch such a program with real data, rather than out of fear of the unknown.
Cybersecurity decisions made under the influence of panic tend to be expensive, poorly designed and unstable. The same decisions, made after careful analysis of trends, work much better. The question isn’t whether we’ll be able to protect ourselves before Anthropic’s next press release. The question is whether we’re building systemic defenses or reacting to hype.
