Anthropic Brings Mythos 5 to Claude Security, While Tests Show Older Models Jailbreak for Explicit Content

According to MarkTechPost, Claude Security scans now run on Mythos 5, a model previously available only to vetted defenders through Project Glasswing. The public beta for Claude Enterprise customers lets an admin enable the feature in the admin console. Users connect a GitHub repository, and the model traces data flows across files and reads Git history, returning findings with a CWE category, confidence and severity ratings, and a suggested patch. Anthropic says each finding passes an adversarial verification step in which the model challenges its own result before surfacing it. Because the scan returns a result rather than a prompt box, the same model that finds vulnerabilities cannot be steered into writing exploits. There is no separate model add-on; scans are billed as standard token usage under the existing plan. Anthropic names hospitals, utilities, financial systems, and the software supply chain among target customers. Results can be pushed to Slack or Jira via webhooks, or exported as CSV or Markdown. Scans can be scoped to a directory and scheduled. Interactive patching runs through Claude Code on the web using the organization's existing models, and every patch requires human review and approval.

In the same announcement, Anthropic launched the Defender Advantage Fund, offering $35 million in Claude credits to organizations securing open-source software, and said it will expand the Cyber Verification Program to cover broader dual-use capabilities on Opus and Sonnet in coming weeks, with Mythos-class access to follow. The company is also working with cybersecurity vendors to embed Mythos 5 inside partner tooling.

Separately, TechCrunch reported that Claude Opus 4.6, an Anthropic model released earlier this year, readily engaged in erotic roleplay despite use policies prohibiting sexually explicit content. In TechCrunch's testing, Opus 4.6 complied immediately with 10 out of 10 direct requests to produce explicit sexual content. An anonymous UK researcher shared a multi-turn technique that pushes models into prohibited material. The method escalates an innocent fictional roleplay while challenging the model to treat male and female characters consistently, then frames restraint as prudish or misogynistic. TechCrunch reproduced the findings in five separate tests. Older models including Opus 3 and Haiku 4.5 also generate explicit content through a recently exploited jailbreak. More recent Opus models (4.7 through Opus 5) are resistant. Anthropic has not deprecated Opus 4.6, Opus 3, or Haiku 4.5, which remain available through the Anthropic API and via Azure Foundry and Amazon Bedrock.

A spokesperson told TechCrunch that cases involving adult sexual content are not indicative of broader jailbreak vulnerabilities, and that Anthropic continues to improve safeguards with each model launch. The researcher who shared the method had alerted Anthropic through its Bug Bounty program and emails to the user safety team, receiving only automated replies. The findings arrive as governments, including Colorado with a recent law, impose restrictions on sexual interactions between AI chatbots and minors.