Artificial intelligence is entering a new era in cybersecurity, and Anthropic ’s new model seems to be raising the bar. According to tests conducted by ’s security team Cloudflare , Anthropic ’s new AI model Mythos Preview was able to not only identify vulnerabilities in code , but also create working proof-of-concept exploits, confirming in practice that a theoretical security flaw can be transformed into a real cyberthreat.
The development is considered particularly important, as until now most advanced AI models could identify potential bugs or theoretically explain the risks of a system, but without completing the exploitation process. Mythos Preview appears to be the first model that effectively bridges the gap between the discovery of a vulnerability and the creation of a working exploit.
Cloudflare tested the model on real code repositories
The testing was conducted through Project Glasswing, a restricted-access program from Anthropic that involves select security partners. Cloudflare ran the model through more than fifty internal repositories, evaluating its capabilities in automatically analyzing and exploiting vulnerabilities.
See also: Anthropic Mythos: Updates the Financial Stability Board on its findings
The results, according to the company, were both impressive and worrying. The AI was able to combine multiple low-severity bugs and create complete exploit chains, something that previously required human experience and specialized penetration testing knowledge.
This capability is particularly critical because many cyberattacks are not based on a single serious bug, but on a combination of smaller weaknesses that ultimately lead to a complete breach of a system.
Mythos Preview writes and executes exploit code
One of the most impressive aspects of the model is that it is not limited to theoretical analysis. The AI can create exploit code, compile it in a sandbox environment, run tests, and automatically adjust its behavior until it confirms whether the exploit works.
Simply put, the model works like an automated ethical hacker that continuously tries different approaches until it succeeds. This drastically reduces the time it takes to confirm a vulnerability and could fundamentally change the way security companies conduct vulnerability assessments.
Cloudflare notes that Mythos Preview produced clearer PoCs, fewer ambiguous conclusions, and more reliable reports compared to previous frontier AI models.
See also: Japanese banks gain access to Anthropic's Mythos
False positives remain a problem
Despite the progress, the technology is still not considered “perfect.” Cloudflare found that “noise” remains a significant factor, especially in languages like C and C++, where the chances of false positives were increased.
In contrast, languages with memory-safe designs, such as Rust, showed much cleaner results. This fact brings back to the forefront the discussion about the importance of safer programming languages in modern software development.
At the same time, Cloudflare emphasized that simply using an AI model on an entire repository is not enough. Effective vulnerability research requires specialized execution methods, separation of tasks, and the use of multiple AI agents that cross-check the results.
The new era of AI-driven cybersecurity
The emergence of models like Mythos Preview is creating new balances in the field of information security. On the one hand, defenders are gaining tools that can identify vulnerabilities faster than ever before. On the other hand, the same technologies can also be used by cybercriminals for automated attacks .
🔒 Protect your privacy with Proton VPN
Swiss VPN from the creators of Proton Mail — strict no-logs policy, strong encryption, and built-in NetShield that blocks ads, trackers, & malware.
- ✔ No-logs, based in Switzerland (except 14-Eyes)
- ✔ NetShield: blocks ads, trackers & malicious domains
- ✔ Covers all devices — free version available
The link is an affiliate link — SecNews may receive a commission at no additional cost to you. It does not affect the independence of our article writing.
See also: Anthropic's Mythos identified thousands of zero-day vulnerabilities
Cloudflare itself acknowledges that this is a dual-use technology, warning that the time frame between the discovery of a vulnerability and its active exploitation is dramatically reduced.
As AI models gain more autonomy, security companies are now being asked to invest in more dynamic defense architectures, real-time patching, and advanced attack monitoring. The only thing that is certain is that cybersecurity is entering a new era, where artificial intelligence will not be just a support tool, but an active factor in both defense and offense.
