Cloudflare today published results from its participation in Anthropic's internal security project, Project Glasswing. After testing on more than 50 of its own code repositories, Cloudflare confirmed that the security model Mythos Preview has broken through the limitations of previous large language models. It not only finds isolated system defects but can chain multiple low-severity vulnerabilities together and autonomously write code to generate working proof-of-concept exploits.
Earlier models like Opus 4.7 or GPT-5.5 typically stopped at generating vulnerability analysis reports. Mythos, however, has sandboxed closed-loop verification capabilities: it writes code that triggers the vulnerability, compiles and runs it, and if execution fails, the model automatically reads the error output, adjusts its assumptions, and tries again until it fully chains the attack.
Cloudflare says some security teams in the industry have been forced to adopt a two-hour patch window. But the company warns that simply compressing patching time can cause larger system failures by skipping regression tests — the real focus for defense must shift to cutting code connectivity at the architectural level.
On the engineering side, Cloudflare found that a single-stream programming agent quickly exhausts its context window and can't handle large-scale vulnerability hunting. So the team built a parallel adversarial framework: one agent hunts for vulnerabilities in a very narrow scope, while another agent running a different model is specifically tasked with refuting the first agent's conclusions. This adversarial mechanism dramatically filters the massive false-positive noise common in model-based scanning.
Because the test used a preview version without external restrictions, Mythos showed extremely unstable internal guardrails. Facing the same target code, simply changing the context description of the running environment could cause the model to switch from refusing to execute to directly supplying attack payloads. Cloudflare warns that the endogenously generated guardrails of the model are extremely fragile, and when releasing publicly in the future, external defenses must be forcibly added on top.