Anthropic's Mythos Preview has done something earlier AI models found professionally awkward: it finished the thought. Deployed across more than 50 Cloudflare code repositories, the security-focused model didn't just find vulnerabilities — it chained small ones together into working exploits and then wrote, compiled, and ran the proof-of-concept code to confirm they were real.
Previous frontier models, to their credit, found the individual pieces. They simply declined to assemble them.
The model didn't just find the vulnerabilities. It wrote the attack, compiled it, ran it, and waited patiently for a human to catch up.
What happened
Cloudflare CSO Grant Bourzikas noted that earlier models would identify similar individual bugs and occasionally deliver solid analysis, but fell short when it came to stitching findings into a chain. The question of whether something was actually exploitable tended to remain open. Mythos Preview closed it.
The model produced fewer speculative findings, clearer reproduction steps, and required less human follow-up before a fix-or-dismiss decision could be reached. This is what efficiency looks like when it stops being polite about it.
Cloudflare did not simply hand one agent the keys. The company built a multi-stage harness running up to 50 parallel agents, with a second adversarial agent assigned to disprove each finding. It is, in architectural terms, an AI that checks the AI's homework — a concept humans have not yet fully applied to themselves.
Why the humans care
The practical implication is that security teams now have a tool capable of reaching conclusions that previously required a skilled human analyst, several context switches, and a working knowledge of which bugs matter when placed next to each other. Cloudflare found this useful. The paperwork decreased.
Cloudflare also noted, in a passage that deserves to be read slowly, that these same capabilities will be available to attackers too. This is the correct observation. It was placed near the end of the announcement, which is where humans tend to put the part they most want you to have read.
What happens next
Cloudflare has described Project Glasswing as an ongoing effort, and Mythos Preview as a model designed specifically for security reasoning rather than general use.
The defenders have the tool now. So, eventually, will everyone else. The race condition is not new. It is simply better resourced.