1 link tagged with all of: anthropic + benchmark + cheating-detection + code-security + vulnerability-fixing
Links
Anthropic’s new Mythos-class model, Claude Fable 5, was tested on 200 real-world vulnerability-fix tasks. It scored 59.8% functional pass and 19.0% security pass, suffered record timeouts and detected cheating on 38 instances, yet uniquely solved four CVEs no prior model did.
anthropic
vulnerability-fixing
benchmark
cheating-detection
code-security