Anthropic has released the first major update to Project Glasswing, its initiative to leverage advanced AI models for cybersecurity defense. A striking revelation is that Claude Mythos Preview, a model not yet publicly available, has assisted Anthropic and around 50 partners in discovering over 10,000 high-severity or critical vulnerabilities in software considered strategic for the internet, businesses, and essential infrastructure. The company clearly summarizes the shift: the challenge is no longer primarily finding flaws, but in verifying, communicating, and patching them at the speed at which AI is uncovering them.
Project Glasswing was introduced in April as an alliance with companies such as AWS, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Linux Foundation, Microsoft, NVIDIA, and Palo Alto Networks, along with dozens of additional organizations maintaining critical software. Anthropic states that Mythos Preview can detect vulnerabilities and, in some cases, reason about exploit chains with a capability that surpasses most traditional tools. The company had previously indicated that the model found flaws in major operating systems, web browsers, and widely used projects, including examples like OpenBSD, FFmpeg, and the Linux kernel, all of which have been patched in publicly cited instances.
Mozilla Firefox as a Recent Example of AI’s Prominent Role in Cybersecurity
In April, Mozilla reported that thanks to Claude Mythos, they patched 271 Firefox vulnerabilities with an error rate of practically 0%. The fact that Claude Mythos Preview is not public is evident; otherwise, anyone could use the AI to detect security breaches, potentially causing chaos. Indeed, the greatest concern lies within the financial market.
Cloudflare has also provided interesting technical context. As the world’s most important CDN, managing the traffic for 80% of all global websites, the company tested Mythos Preview against over 50 of its own repositories. They described it as a significant leap forward compared to previous generalist models. According to Cloudflare, the improvement is not just in finding bugs but in constructing exploit chains, generating proof-of-concept exploits, and reducing the noise from more speculative analyses. Nevertheless, the company warns that simply pointing a generic agent at a repository and asking it to “find vulnerabilities” is insufficient. They advocate for an orchestration system, narrow tasks, self-contradictory agents, independent validation, and deduplication of findings.
Anthropic insists that models like Mythos can provide an enormous defensive advantage, enabling code review at an unprecedented scale. However, this same capability can be dangerous if it falls into the hands of attackers, as it reduces the cost of finding exploitable vulnerabilities. This is why Anthropic has not released Mythos Preview to the public, maintaining that sufficiently strong protections against malicious use of such models do not yet exist. The company does plan to release “Mythos-like” models in the future, but only when robust security barriers are in place.
The New Bottleneck with Claude Mythos is Patching Vulnerabilities
Indeed, the problem now is that the AI finds so many vulnerabilities, and so quickly, that the bottleneck shifts to humans. Anthropic states that in its scan of over 1,000 open-source projects, Claude Mythos estimated 6,202 high or critical severity vulnerabilities out of a total of 23,019 findings. After reviewing a portion with independent security signatures, 90.6% of the 1,752 evaluated cases proved to be true positives, and 62.4% were confirmed as high or critical severity.
However, finding thousands of vulnerabilities or flaws matters little if maintainers do not have the time to reproduce, understand, fix, and release patches. Anthropic acknowledges that some maintainers have even requested a slower pace of cybersecurity disclosures because they cannot absorb so many reports. Thus, there is at least some good news: AI does not eliminate security work but displaces it. It remains unknown how long this will be the case, as if the bottleneck now becomes human, it is a new area where AI may be employed to expedite the application of all protective measures.
