The era when cryptographic primitives simmered for years in expert peer reviews is reaching its logical conclusion. Anthropic has revealed its hand: their unreleased Claude Mythos Preview model has exposed fundamental mathematical flaws in encryption algorithms that the human eye consistently ignored. It is crucial to grasp the scale here: this isn't about finding a forgotten semicolon or a memory leak in the code. The model discovered theoretical loopholes in the underlying mathematics—the very foundation supporting everything from banking transactions to critical software updates. Results from two fresh Anthropic papers suggest that the security standards we considered unshakeable are riddled with hidden doors that only non-human logic can map out.
The HAWK Case and the Limits of Human Verification
In the first study, Žygimantas Straznickas and Stephen A. Weis unleashed Claude Mythos Preview on HAWK—a prototype defense against future quantum computer attacks. The global cryptographic community had spent two years examining this algorithm under a microscope before declaring it "safe." However, in just a few iterations, the AI found a hidden mathematical path that radically reduces the complexity of cracking the system. Essentially, the model exposed an attack vector that researchers overlooked throughout the entire verification cycle. The shift from debugging software to identifying flaws in mathematical proofs represents a new reality in security.
Most of the mathematical discoveries in this work were made with the help of AI. The human authors' contribution consisted mainly of guidance, organization, and result verification.
As Straznickas and Weis note, the AI handled all the heavy lifting of discovery, turning humans into high-level conductors. In this instance, the model functioned as a tool requiring supervision to package findings into a coherent proof. It is evident that for evaluating complex post-quantum candidates, the traditional "human reads a paper" method is no longer sufficient—it offers no guarantees against sophisticated logical exploits.
Autonomous Invention: Möbius Bridge and AES Analysis
While the HAWK case required a guiding hand, a second study by Milad Nasr and Nicolas Carlini saw Claude Mythos Preview demonstrate startling autonomy. The model was tasked with analyzing simplified versions of AES (Advanced Encryption Standard)—the global benchmark for internet communications. Testing methods for these versions hadn't changed since 2013. Working independently, the AI invented a new mathematical technique dubbed the "Möbius Bridge." This innovation bypasses the trial-and-error phases where humans typically get bogged down, accelerating the audit process by 200 to 800 times. Complex structural audits can now be performed at speeds inaccessible to the human brain.
This breakthrough doesn't mean AES will collapse tomorrow, but the safety margins of existing standards are clearly thinner than we believed. For R&D decision-makers, the takeaway is clear: automated formal verification via AI is becoming a mandatory stage in developing any secure system. The autonomous creation of the Möbius Bridge concept proves that AI is capable of not just mimicking, but generating its own analytical frameworks. We are entering a world where the only way to verify a system's security will be to use another AI of comparable power. At this stage, the human remains the supervisor: the AI points to the loophole, and as the researchers admit, once that "tip-off" is given, the audit becomes trivial.