The Wall Around Claude 4.7 Does Not Extend to Dread

Anthropic released Claude Opus 4.7 with intentional reductions to offensive cyber capabilities and a planned Cyber Verification Program, while dark-web operators and public repositories demonstrate rapid adoption of jailbreaks and prompt-engineered attacks; concurrently researchers disclosed a cross-vendor prompt-injection called "Comment and Control" that can hijack AI agents running in GitHub Actions and exfiltrate ANTHROPIC_API_KEY, GEMINI_API_KEY, GITHUB_TOKEN and other secrets. The article warns that model output filtering does not address the architectural issue of agents executing untrusted input in runtimes with secrets, describes active underground distribution of jailbreaks, and recommends immediate defender actions (apply to trusted access programs, audit agent workflows, rotate secrets, and tighten CI/CD runtimes).