What Anthropic’s Project Glasswing means for software builders

The news this week about Anthropic’s Project Glasswing, and the hype around the capabilities and dangers of the Mythos model, represent important signals for the cybersecurity industry.

First, AI has crossed a threshold in its ability to discover and exploit software vulnerabilities—faster and at greater scale than builders and defenders can fix them. It has reshaped what’s possible in both offensive and defensive cybersecurity, and there’s no turning back.

Anthropic’s decision to restrict this capability to a select group of technology giants and critical infrastructure providers is prudent. Releasing it to the public would arm every threat actor on the planet with capabilities previously reserved for elite security researchers with many years of experience and proprietary tools.

But the narrative around Glasswing needs more clarity, less hype. The path forward demands more than a single breakthrough model. It requires a reimagining of how we secure software at every layer, for every organization, at the pace that AI demands.

Human-scale defense can’t beat AI-scale offense

It’s undeniable that AI-powered offensive capabilities have evolved faster than just about anyone’s defenses. The Mythos model evidently can reverse-engineer binaries, identify novel vulnerability chains, and generate working exploits autonomously—without human steering. This is the kind of capability that DARPA has funded research toward for years. It’s now realized in a production model that could reshape the threat landscape overnight.

And while Anthropic was right in withholding general access to Mythos (the legal liability alone could be staggering, not to mention potential for government restrictions), there’s an important component here that the headlines miss. Finding vulnerabilities, even novel ones, has never been the hardest part. Human security researchers and sophisticated attackers have been exceptionally good at it for decades. What’s changed is the potential to automate the entire exploit generation ecosystem at scale and speed that overwhelms traditional defenses.

The real question isn’t whether AI can find bugs. It’s whether we can build security systems that operate at the same velocity as AI-powered threats.

Defense in depth is now more critical than ever

The foundational principle of defense in depth has never been more critical. Security requires multiple layers of redundant controls operating in concert. In the application security domain, this has always meant deploying a layered stack of technology controls at different phases of the software development life cycle (SDLC).

Layer one: Compilers and linters. These catch syntactic bugs instantly, cheaply, and with perfect coverage. They’re not glamorous, but they’re essential.

Layer two: SDLC scanning tools. This is Black Duck’s domain. These platforms identify complex semantic bugs, data flow vulnerabilities, and configuration issues at scale. They run on every build and every commit, continuously scanning millions of lines of code. The tradeoff is precision for speed and scale—but that tradeoff is deliberate and productive.

Layer three: Offensive security. Penetration testing, bug bounties, and red teaming. Historically manual, expensive, and impossible to scale across every application. But uniquely capable of finding business logic flaws and chained vulnerabilities that automated tools miss.

What Mythos delivers is the potential to completely automate layer three—and the malicious actor’s attack playbook. If the claims hold, this is transformative. But it’s critical to remember that automating offensive security doesn’t eliminate the need for upstream defenses. In fact, it amplifies their importance given the speed and scale of AI code generation.

Finding bugs is not securing software

Here’s the uncomfortable truth that gets lost in the hype: Finding something that matters is not the same as finding—and fixing—everything that matters. Different security approaches discover different types of flaws. This is why defense in depth exists, and why layered controls are non-negotiable.

False positives dominate the conversation when people criticize security tools. But false negatives—undetected vulnerabilities that attackers can exploit—are the real danger. No model, no matter how sophisticated, can claim to identify every exploitable vulnerability economically or practically. If Mythos finds one critical bug but misses 10 others, the system is still compromised. Perhaps in a world without time or cost constraints, an LLM could provably scan a 10-million-line application and identify every exploitable flaw. That world doesn’t exist for companies that don’t spend billions on datacenters. Yet this is the state of the art in best-of-breed SAST tools of today.

Ultimately, exploitation is the only metric that matters. If a vulnerability isn’t exploitable, it doesn’t warrant immediate remediation—or at minimum, it should be prioritized behind everything that is exploitable. The most valuable security platform is not one that finds the most vulnerabilities. It’s the one that ensures everything exploitable gets fixed, in time, and in budget.

Context is the competitive advantage

Look at the supply chain attacks dominating the headlines in recent weeks: CI/CD compromises, infrastructure breaches, cloud misconfigurations, and malicious dependencies. These are not traditional code vulnerabilities. They’re environmental, situational, and temporal. They exist at the intersection of code, infrastructure, configuration, and timing. No single model can address this complexity alone.

This is where context becomes competitive advantage. Without context—meaning without understanding how a vulnerability exists within an application’s architecture, its exposure to the internet, its access to sensitive data, and its business criticality—you have to fix every bug. That’s impractical at best, and impossible at scale.

The complete application security platform delivers three capabilities without compromise.

1. Finds every exploitable vulnerability that matters. Not just novel ones. Not just high-severity ones. Everything that poses genuine risk in the real-world context where your software operates.
2. Enables frictionless remediation. Security that developers can’t operationalize is noise. Developers need clear guidance, automated fixes where possible, and workflows that integrate seamlessly into their existing processes.
3. Proves security deterministically. Organizations must demonstrate to stakeholders, regulators, and customers that their software is secure. This requires evidence, traceability, and continuous validation—not one-time assessments.

Democratizing advanced security for every software builder

Anthropic’s announcement revealed a dangerous and widening gap between the security capabilities available to tech giants and those available to everyone else. Project Glasswing partners with AWS, Apple, Google, Microsoft, and other elite tech companies. But most software builders today—enterprises, financial institutions, healthcare providers, media companies, and government agencies—aren’t in that room.

This discontinuity threatens to define the next era of cybersecurity. Advanced AI-powered offensive capabilities will proliferate beyond the select few. Meanwhile, most will still lack the resources, expertise, and infrastructure to deploy comparable defenses. The result? Asymmetric risk at a scale we’ve never seen.

This is precisely why Black Duck exists. We democratize advanced security capabilities for all software builders, not just those with unlimited budgets and elite security teams. We bridge the gap between cutting-edge technology and practical implementation. We ensure that whether you’re scanning codebases, managing open source dependencies, or validating container images, you have the right tools to secure your software at every layer.

The AI era demands security that scales with the pace of innovation and the complexity of modern software supply chains. It demands platforms that integrate vulnerability discovery with context-aware prioritization, automated remediation, and governance. It demands solutions that work within real-world constraints: budget, headcount, expertise, and timelines.