Cloud and security company Cloudflare disclosed results from deploying Anthropic's Mythos preview on its own infrastructure.

Cloudflare is among participants in Project Glasswing, under which Anthropic is providing Mythos on a limited basis to some companies and institutions. Cloudflare applied Mythos to more than 50 internal repositories and shared in detail on its website what it found and what limits it encountered.

Anthropic is providing Mythos only to some users through Glasswing because it can find software vulnerabilities so well that it could be abused for cybersecurity purposes.

Cloudflare assessed Mythos, compared with previous general-purpose frontier models, as "a tool on a different level, not a simple improvement."

It highlighted 2 capabilities. One is building attack chains.

According to the company, real attacks do not use just 1 bug. It is common to create an exploit that works by linking multiple small attack elements. Mythos can combine these elements to generate code that proves a real attack is possible.

Cloudflare said the reasoning process looks like the work of a skilled security researcher rather than an automated scanner.

The other is proof generation. Finding a bug and proving it can be exploited are different.

When Mythos finds a bug in code, it directly writes and runs test code to verify whether it can be exploited. If exploitation is confirmed, it judges it a real threat. If it is not confirmed, it changes its approach and tries again.

Cloudflare said other frontier models also found a significant number of bugs in the same environment, but often stopped after finding and describing individual bugs.

It stressed that what is different about Mythos is that it links bugs that are low severity on their own into a single serious exploit.

Cloudflare pointed out that Mythos sometimes refused legitimate requests for defensive vulnerability analysis. It said that framing the same task differently sometimes produced the opposite result.

Cloudflare warned that as AI finds vulnerabilities quickly, "many security teams are aiming to patch within 2 hours of vulnerability disclosure, but focusing on speed alone can be risky," adding that "skipping regression testing can create problems more serious than the original bug." It stressed that more important than speed is an architecture that makes it difficult for attackers to exploit vulnerabilities even if bugs exist.