IBM plans to build Anthropic's Claude Mythos into Concert Protect, with a target of early 2027. The goal is to let customers point the model at their own code.
As described, the feature does three things. It scans code for flaws nobody has catalogued yet, tests which of those flaws an attacker could actually exploit, and produces fixes for developers to approve. Mythos itself remains limited to vetted organisations, so this is a plan to widen access through an IBM product rather than a general release.
The middle step is the one to watch. Scanners are good at producing long lists, and a list of unconfirmed findings mostly hands the work to a human. A tool that separates exploitable flaws from theoretical ones addresses that problem, assuming it works as described. IBM has offered no performance figures, and the target date is still a target.
The design also keeps developers in the approval seat. The model proposes the fix, and a person signs off before anything changes.
The open question is access. Anthropic restricts Mythos to vetted organisations, and the announcement does not say whether IBM customers would need to clear a similar bar, or how much of the model's capability the product would expose.