Anthropic Expands AI Testing Program for Cyber Teams After Finding 129,000 Flaws
Anthropic is broadening access to its AI safety testing initiative, citing the discovery of over 129,000 verified vulnerabilities by vetted researchers in recent months.
Key points
- Anthropic expanded its Project Glasswing program to allow more cybersecurity professionals to test AI models with reduced safeguards.
- The initiative uncovered at least 129,000 verified software vulnerabilities between April and July 2026.
- The expanded access includes testing with blocking classifiers removed to identify deeper security issues.
Anthropic announced on Tuesday that it is expanding Project Glasswing, a program that grants vetted cybersecurity professionals access to its advanced artificial intelligence models with reduced safety measures. The company stated that this expansion follows the initiative's identification of at least 129,000 verified software vulnerabilities between April and July 2026, according to The Hacker News.
The expanded program allows qualified security researchers to test Anthropic’s Claude models with fewer restrictions. Specifically, participants can evaluate the AI systems with reduced safeguards and blocking classifiers disabled. This approach aims to uncover potential security weaknesses that might remain hidden under standard operating conditions. Anthropic indicated that the previous phase of the initiative had already yielded significant results, prompting the decision to broaden participation among vetted teams.
Context and Background
Project Glasswing serves as a structured channel for external security experts to probe Anthropic’s AI infrastructure. By temporarily lowering certain safety barriers, the company enables testers to identify vulnerabilities that could be exploited in real-world scenarios. The announcement highlights the scale of findings from the initial period, with the 129,000 verified flaws representing a substantial volume of security issues addressed through this collaborative effort.
The timeline of the reported findings covers a four-month window from April to July 2026. This period allowed researchers to systematically evaluate the models and report their discoveries to Anthropic. The company’s decision to expand the program suggests a continued reliance on external expertise to maintain the security integrity of its AI systems.
Who Is Affected
The primary beneficiaries of this expansion are vetted cybersecurity professionals and teams who have been approved to participate in Project Glasswing. These individuals gain access to testing environments that differ significantly from standard public-facing deployments. For the broader technology community, the announcement signals a shift in how AI developers are engaging with security research, emphasizing proactive vulnerability discovery over reactive patching.
Anthropic’s users are indirectly affected by these efforts, as the identification and remediation of vulnerabilities contribute to the overall security posture of the Claude models. However, the specific details of the vulnerabilities or the nature of the fixes applied are not disclosed in the initial announcement. The focus remains on the process of discovery and the collaboration with the security research community.
What should you do?
Security professionals interested in participating in similar initiatives should monitor official channels for application requirements and vetting processes. Organizations using AI models should stay informed about vendor-led security programs and consider engaging with their providers to understand how external testing impacts product security.
What to do and how to stay safe: Anthropic
- Review vendor announcements regarding security testing programs to understand how external research impacts product integrity.
- Ensure that internal security teams are aware of new testing initiatives that may reveal vulnerabilities in third-party AI services.
- Monitor for updates from AI providers on how discovered vulnerabilities are being addressed and communicated.
- Evaluate whether your organization’s use of AI models requires additional oversight given the volume of vulnerabilities found in similar systems.
Step-by-step guide: Prompt Injection Warning Signs: How to Spot AI Manipulation
General security guidance from the Threat Times newsroom. It is not confirmed advice from the organisations named in this story.
Frequently asked questions
How many vulnerabilities did Project Glasswing find?
The initiative uncovered at least 129,000 verified software vulnerabilities between April and July 2026.
Who can participate in the expanded program?
Vetted cybersecurity professionals and teams are eligible to test the AI models with reduced safeguards.
What changes are being made to the testing environment?
Participants will have access to models with reduced safeguards and blocking classifiers disabled.
