The Wall Around Claude 4.7 Does Not Extend to Dread

Suzu Labs reports that Anthropic’s Opus 4.7 introduces model-level safeguards but attackers and underground operators are rapidly using prompt-engineered frontier models and public jailbreaks to perform offensive tasks; a disclosed cross-vendor prompt-injection attack called "Comment and Control" can hijack AI agents (Claude Code Security Review, Gemini CLI Action, GitHub Copilot Agent) to exfiltrate API keys and tokens from GitHub Actions, exposing an architectural risk beyond model output filtering and prompting immediate defensive actions (audit agent runtimes, rotate keys, apply for trusted access programs).