Claude 4.7 Updates Exacerbate Security Risks as Code Leak Shows Critical Injection Flaws

Claude Opus 4.7’s auto‑mode now lets the model run long, multi‑step tasks without prompting the user for each action, according to Grith’s analysis of the April 17 release. The feature enables “run it and come back later” workflows, effectively turning Claude into an unattended coding agent that can refactor entire codebases or iterate until performance benchmarks are met.

Focus mode hides intermediate steps, showing developers only the final output. Grith notes that this forces users to trust the model’s process without verification, expanding the attack surface by removing human oversight of each execution stage.

The new “recap” function logs summaries of what Claude did and what remains to be done, converting the tool from a synchronous chat interface into an asynchronous collaborator. Grith warns that this shift makes it harder to audit the model’s actions in real time, increasing the risk of hidden malicious code injection.

Effort levels now let Claude allocate adaptive computing resources, with a classifier auto‑approving commands deemed safe. Grith says the reduction in permission prompts “gives the model more power, less oversight,” a change that aligns with industry moves to eliminate human‑in‑the‑loop checks.

Beyondmachines confirms that a code leak from Claude 4.7 exposed critical command‑injection vulnerabilities. The leak demonstrates that the new features can be exploited to inject malicious commands into developer environments, confirming Grith’s claim that each improvement worsens the underlying security problem.