Anthropic expands cyber testing access as Project Glasswing finds 129,000 vulnerabilities

Key takeaways

• Anthropic is giving vetted security researchers access to Claude with reduced guardrails to enable deeper vulnerability testing.
• Project Glasswing found 129,000 verified software vulnerabilities in four months using AI-assisted methods.
• The program expansion shows AI companies moving toward tiered access models instead of blanket restrictions.
• The pace of AI-driven vulnerability discovery may outstrip many organizations' ability to patch quickly.
• Businesses should prioritize faster patching cycles and consider AI-assisted security testing for their own systems.

What happened

Anthropic announced on October 6 (https://thehackernews.com/2026/10/anthropic-expands-claude-access-for.html) that it is expanding a program allowing vetted cybersecurity professionals to test Claude with reduced safeguards and blocking classifiers. At the same time, the company reported that Project Glasswing, its own security research initiative, uncovered at least 129,000 verified software vulnerabilities between April and July 2026.

This is a notable shift. Most AI companies lock down their models tightly to prevent misuse. Anthropic is doing the opposite for a select group: loosening the guardrails so security researchers can push Claude harder and test what it can find.

The timing is deliberate. By publishing results from Glasswing alongside the program expansion, Anthropic is making a case that frontier AI models are powerful tools for defenders, not just theoretical risks in the hands of attackers.

Why this is an issue

AI models capable of finding vulnerabilities at scale create a dilemma. If they can help defenders discover flaws, they can also help attackers exploit them. The question is not whether these capabilities exist but who gets to use them and under what conditions.

Anthropic's approach is to carefully vet who gets access. The expanded program is not open to everyone. Security professionals must pass a screening process before they can work with Claude under reduced restrictions. That vetting process is the control layer.

But vetting is not perfect. It introduces a judgment call: who counts as trustworthy? And what happens when a vetted researcher's credentials are compromised or when a bad actor fakes their way through the process? These are known risks in any credentialing system.

The 129,000 vulnerabilities number also raises a practical concern. That volume of findings in four months is far more than most organizations can patch quickly. If AI can find flaws faster than humans can fix them, the result may be a growing backlog of known but unaddressed weaknesses. Defenders gain visibility, but they also inherit a workload problem.

What you can do about it

If you run software or manage systems, assume that AI-assisted vulnerability discovery is already in play. Attackers and defenders both have access to models that can probe code and configurations. Your posture should reflect that reality.

First, prioritize patching cadence. If vulnerabilities are being discovered faster, your update cycle needs to match. Review how quickly your team applies security patches after vendors release them. If it takes weeks, tighten that window.

Second, if you develop software in-house or work with custom applications, consider using AI-assisted security testing yourself. Tools that incorporate language models for code review and vulnerability scanning are increasingly accessible. You do not need to be a security expert to run automated scans, but you do need someone who can interpret results and prioritize fixes.

Third, review your vendor security practices. Ask your software providers how they discover and remediate vulnerabilities. If they are not using modern tooling, including AI-assisted methods, they may be falling behind the pace of discovery.

Finally, if you handle sensitive data, audit your exposure. AI can find flaws in public repositories, open-source dependencies, and poorly secured APIs. Make sure your credentials, keys, and internal configurations are not leaking in places a model could scrape.

Where this is heading

Expect more AI companies to adopt tiered access models. The binary choice between locked-down and fully open is giving way to a middle ground: restricted access for vetted users who need the model to do security work that would otherwise trigger blocks.

This will likely extend beyond cybersecurity. Researchers studying disinformation, content moderation, and other adversarial use cases will push for similar carve-outs. The challenge for AI companies will be scaling the vetting process without losing control.

We will also see a scaling race in vulnerability discovery. If one model can find 129,000 flaws in four months, the next generation will find more. The bottleneck will shift from discovery to remediation. Organizations that cannot keep up with patching will become soft targets, not because they lack visibility but because they lack bandwidth.

Our read is that this moves security work from art to assembly line. Finding vulnerabilities has always required skill and intuition. If AI can automate much of that, the scarce resource becomes the human judgment needed to prioritize, patch, and architect systems that are resilient even when flaws exist. The work does not disappear. It changes shape.