Anthropic is quietly testing what could be one of its most powerful AI systems yet. And instead of rushing it to market, the company is doing something different.
It’s letting a small group of the world’s largest tech companies try to break it first.
What is happening
Anthropic has launched an initiative called Project Glasswing, giving early access of its unreleased AI model Mythos to companies like:
- Apple
- Amazon
- Microsoft
- Cisco Systems
The goal is simple but important:
Let defenders find vulnerabilities before attackers do.
These companies are using Mythos to scan their systems, identify weaknesses, and share insights across the industry.
Why this matters
The next generation of AI models is not just better at writing code or answering questions.
They are getting exceptionally good at finding flaws.
That creates a double-edged situation:
- Positive: AI can help companies detect vulnerabilities faster than ever
- Risk: The same capability can be used by hackers to exploit systems at scale
This is exactly the concern Anthropic is trying to address early.
Instead of reacting after misuse happens, it is proactively stress-testing the model in real-world environments.
What Mythos is already capable of
Even in early testing, Mythos has shown why companies are taking this seriously.
- It identified a 27-year-old bug in widely used internet infrastructure
- It found a 16-year-old vulnerability in video game software
- That vulnerability had been scanned millions of times by traditional tools without detection
This is the key shift:
AI is moving from assisting developers to outperforming traditional security systems.
Why Anthropic is not releasing it yet
Unlike the typical AI race where companies push models out quickly, Anthropic is taking a more cautious approach.
- Mythos is not being released to the public yet
- Feedback from Project Glasswing will be used to build guardrails and safety layers
- The company is also in discussions with US government bodies on security implications
This reflects a broader realization across the industry:
More powerful AI does not just scale benefits. It also scales risk.
This is not just Anthropic’s problem
Even competitors like OpenAI have acknowledged similar risks and launched programs to give defensive teams early access to advanced models.
What’s emerging is a new pattern:
- AI companies are no longer just building tools
- They are also responsible for how those tools reshape cybersecurity dynamics
And importantly, this is becoming a shared industry effort, not a competitive one.
The bigger picture
Project Glasswing signals something deeper about where AI is headed.
- AI is becoming a core part of cybersecurity infrastructure
- The line between builder and breaker is getting thinner
- Companies are preparing for a future where AI vs AI becomes the norm in cyber defense
In simple terms:
The same technology that can secure systems can also be used to attack them.
And whoever understands it first has the advantage.
What investors and builders should watch
- AI-led cybersecurity could become a massive category on its own
- Companies that integrate AI into defense systems early may have a structural edge
- Collaboration between competitors may increase in areas involving systemic risk
Bottom line
Anthropic is not just testing a new AI model.
It is testing a new approach to releasing powerful technology.
Build first. Break it internally. Strengthen it. Then scale it.
If this becomes the norm, the AI race may shift from who launches first to who deploys safely at scale.
And that could define the next phase of the industry.