4/10/26 Anthropic is holding back a new model, Claude Mythos Preview, because of its potential cybersecurity impact. Instead of public release, access is limited to a consortium of major tech players under Project Glasswing to identify and patch vulnerabilities across critical systems.
Withholding the model at this time is because it so good at discovering vulnerabilities, including flaws missed for decades by human researchers and automated tools – and these vulnerabilities can be exploited by bad actors and cause harm to organizations. What once required elite expertise and months of effort can now happen in minutes.
For security leaders, this means that assumptions about “secure enough” systems no longer hold. If AI can systematically surface hidden weaknesses across global infrastructure, the question becomes not just how we defend — but whether large portions of existing codebases must be revalidated or rewritten.