Cloudflare has released a report, signed by its Chief Security Officer, Grant Bourzikas, detailing Project Glasswing—a collaborative experiment using the Mythos Preview model, a cutting-edge cybersecurity tool developed by Anthropic. The model analyzed over fifty of Cloudflare’s internal repositories autonomously, constructing comprehensive operational chains and creating functional proofs of concept, whereas previous iterations only managed to identify bugs. The report also documents inconsistencies in the model’s refusals and urges the community to ensure that any publicly available version incorporates enhanced safeguards.
The journey began months ago when Cloudflare, operating one of the internet’s most exposed networks, sought to leverage language models to identify vulnerabilities in its own code before cyber attackers could exploit them. For over a year, the company has been testing cybersecurity-focused models, as well as more general frontier models, all of which exhibited a common limitation: while they could identify interesting bugs and accurately describe their potential impact, they fell short at delivering actionable exploits, leaving the operational chain incomplete and the question of exploitability open to speculation.
Weeks prior to the report’s publication, Anthropic invited Cloudflare to test a Preview version of Mythos, designed specifically for cybersecurity tasks. This version lacked the additional safety measures present in generally available models such as Opus 4.7 or GPT-5.5, and Bourzikas accepted. His team pointed Mythos at fifty internal repositories, ranging from the execution engine and edge data path to the protocol stack and control plane, along with open-source projects that Cloudflare relies on.
Two Surprising Technical Autonomy Findings
The first surprise emerged quickly: Mythos did not just flag isolated bugs; it connected them seamlessly. A successful attack rarely exploits a single vulnerability; it typically combines multiple primitives, such as transforming a use-after-free flaw into arbitrary read-write capabilities, redirecting control flow, or executing a return-oriented programming (ROP) chain to assume control of a system. Where previous models fell short in stitching these elements together, Mythos completed the task. Its reasoning during this process resembled that of a seasoned researcher rather than that of an automated scanner.
The second, even more significant surprise was that Mythos did not just reason about the exploitability of a vulnerability; it executed code to demonstrate it. The model would write code that triggered the suspected defect, compile it in a sandbox, and run it. If the program produced the expected outcome, that served as conclusive proof; if not, the model would analyze the failure, refine its hypothesis, and retry. This self-validation loop distinguishes Mythos markedly from earlier generations, bridging the gap between a suspected vulnerability and proof of concept.
Unexpected Refusals and Warnings to Developers
Throughout the experimentation process, unexpected incidents were observed. Even in this version lacking additional safeguards, Mythos sometimes refused certain requests. In one instance, it declined to search for vulnerabilities in a project but accepted the same query after an unrelated change to the environment—without any alteration to the code analyzed. In another case, it detected and confirmed several serious memory bugs but refused to write an exploit demonstration. A slightly rephrased request yielded a different response.
Bourzikas draws a clear conclusion for Anthropic and the wider community of model developers. These emerging refusals are real but inconsistent enough not to serve as a comprehensive safety boundary. Any future cybersecurity frontier model made publicly available must include additional safeguards beyond this fundamental behavior. The warning, articulated by Cloudflare’s security chief following the contractual experiment with Anthropic, carries significant weight—more so than an independent analyst’s report.
A Shift Towards Harnessing and Architectural Conclusions
With these capabilities confirmed, Cloudflare encountered the same limitations as with previous models. When focused solely on a large repository, Mythos overloaded its context before covering the relevant surface area. The team shifted strategies, abandoning attempts to make the model undertake complicated tasks autonomously. Instead, they constructed an eight-step harness (Recon, Hunt, Validate, Gapfill, Dedupe, Trace, Feedback, Report) that assigns narrow, parallel, and verifiable tasks to around fifty competing hunters.
The report concludes with a cautionary note to Bourzikas’s peers. Several security leads consulted by Cloudflare now operate under a two-hour service-level agreement (SLA) for patching vulnerabilities from the moment a Common Vulnerability and Exposure (CVE) is published. However, Bourzikas emphasizes that this response is insufficient. Speeding up the patching process does not address the underlying issues within the pipeline that creates the patches. The real solution lies in architectural improvements that make exploitation difficult, even when a bug exists.



