Anthropic’s New AI Model Can Identify More Software Bugs Than Ever. Microsoft Is Struggling to Fix Them Fast Enough.

Anthropic’s Mythos AI Model Detects Software Vulnerabilities; Microsoft and Other Companies Address Identified Issues

On an afternoon in mid-May, dozens of Microsoft engineers and their managers gathered online and in a conference room at the company’s Redmond, Washington, headquarters to discuss Project Glasswing.

The project involves access granted by Anthropic to a group of organizations that develop widely used software. Anthropic developed the AI model Mythos to identify vulnerabilities in code. The participating organizations include Microsoft along with 11 other launch partners such as AWS, Apple, and Google. The access was provided for the purpose of locating and addressing weaknesses before external actors could do so.

During the meeting, one engineer asked whether the version of the model in use, Claude Mythos Preview, was performing as Anthropic had indicated it would. A manager responded affirmatively, according to a recording of the session reviewed by ProPublica. The manager stated that the model was identifying vulnerabilities at a rate that exceeded the pace of remediation efforts at that time, describing the situation as requiring accelerated work to address the findings.

A slide presented during the meeting indicated that in April the model had identified 90 vulnerabilities rated critical and 141 rated important in SharePoint. Additional findings were reported in the first half of May. Engineering manager Hans Andersen urged teams to prioritize work on the April findings within the remaining access period, which was scheduled to conclude on May 31.

Participants on the call discussed the implications of the access period ending, with one noting that subsequent availability of similar capabilities could allow others to identify the same issues.

Project Glasswing was publicly referenced in April. At that time, national security analysts noted a period during which organizations with access could address findings before comparable capabilities became available more broadly. In late June, the Five Eyes alliance issued a joint statement indicating that this period could close within months. Internal Microsoft documents reviewed by ProPublica show that the volume of findings from Mythos had already required prioritization of higher-severity items.

Microsoft has directed initial remediation efforts toward vulnerabilities classified as critical or important, consistent with its standard triage process. The documents indicate plans to address moderate-severity items at a later stage and do not reference low-severity items. The company has described its triage criteria as including factors such as exploitability and potential impact on users.

Vinh Nguyen, a senior technical adviser to Anthropic and senior fellow for AI at the Council on Foreign Relations who previously served as chief AI officer at the National Security Agency, stated that chaining of multiple lower-severity vulnerabilities can produce higher overall risk. Nguyen indicated that existing triage approaches may assign insufficient weight to such combinations. Microsoft has stated that chaining has been incorporated into its vulnerability assessment processes for an extended period and that it continues to evaluate whether adjustments to severity classifications are warranted as AI-assisted discovery methods evolve.

Microsoft declined to provide specific figures on the number of Mythos-identified vulnerabilities that have been remediated since the May meeting. Anthropic also declined to comment.

The internal presentation projected that work on SharePoint findings would continue for months, beginning with critical items and moving to important items in August before addressing approximately 300 moderate items. Documents reviewed by ProPublica indicate that Mythos had identified hundreds of critical or important vulnerabilities across products including Microsoft 365, Teams, and Copilot since access began earlier in the year, with most unaddressed as of mid-May. Andersen described the identified issues as exploitable in many cases.

It remains undetermined whether any specific Mythos finding has been exploited. Separate reports have documented the use of AI tools by malicious actors to automate aspects of vulnerability discovery and exploitation.

Microsoft’s monthly Patch Tuesday releases provide one measure of remediation volume. The June release addressed more than 200 vulnerabilities, described by some observers as a record at the time. The July 14 release addressed more than 600 vulnerabilities, of which seven were rated low or moderate severity and one was under active exploitation, according to Dustin Childs of the Zero Day Initiative. Microsoft has stated that the overall volume of identified vulnerabilities is not expected to plateau in the near term and that it has increased investment in both personnel and AI-supported triage processes.

Nguyen has suggested that organizations may need to expand resources dedicated to remediation across the full range of severity levels rather than focusing exclusively on higher-rated items. Microsoft has indicated that it is reassessing prior classifications of lower-severity items in light of AI discovery capabilities and that similar evaluations are occurring across the industry.

Microsoft products are frequent targets for malicious activity due to their widespread deployment. Some products incorporate code developed years earlier, which can contain unaddressed issues. The remediation challenge extends beyond Microsoft to other software developers and to open-source projects that rely on volunteer maintenance. J. Michael Daniel, former cybersecurity adviser to President Barack Obama and president of the Cyber Threat Alliance, stated that accumulated technical debt across the sector requires attention. Data scientist Ben Edwards noted that vulnerability volumes were already substantial prior to widespread AI use and have increased further.

Microsoft’s Security Response Center has handled large volumes of reports in prior years. Former employees have described security work as a cost center relative to product development. Microsoft has stated that it does not discuss internal staffing levels but has directed recent investments toward customer security.

The experience at Microsoft reflects conditions reported by multiple organizations participating in Project Glasswing. The scale of findings from Mythos has prompted parallel review processes at other partner companies, each applying their own triage frameworks to the results.