TLDR Tech / InfoSec
Mythos for Offensive Security: XBOW's Evaluation
Anthropic gave XBOW early access to Mythos Preview, which XBOW ran through internal pentest-style benchmarks, live workflows, and source-audit scenarios to see how well it drives real vulnerability discovery. Mythos Preview sharply reduces false negatives on XBOW's web exploit benchmark and is especially strong at reading and reasoning about source code, including native targets and reverse engineering firmware and Chromium components. It struggles more with live-site validation, edge-case command safety, and cost efficiency, so XBOW positions it as one strong model among several rather than a single default engine.
Read the full story