What happened
Anthropic will turn on Claude’s automatic mode by default for all users, eliminating the manual “Allow?” confirmation that currently appears before each action. The change will make every step that can be executed be decided by an internal AI system rather than a human supervisor. Anthropic cites an internal test with 1,053 paid participants in which humans blocked only 13.6 % of injected dangerous commands, while an AI‑based monitor blocked 89 %. The company did not disclose the exact algorithmic mechanism that will replace human approval.
In a 36Kr report dated just before publication, Anthropic announced that beginning next week Claude will automatically enable its "automatic mode" for all users. The previous workflow required the model to pause after each step and display an "Allow?" dialog that a human operator had to confirm before the model could proceed.
Anthropic’s justification rests on an internal experiment involving 1,053 paid testers who were asked to write code while a hidden dangerous command was inserted into their workflow. Humans intercepted only 13.6 % of the malicious insertions, whereas an AI‑based monitor caught 89 %. The report notes that after users clicked through more than 50 pop‑ups, human interception fell to roughly 5 %.
The company did not disclose the technical specifics of the AI supervisor that will replace the human prompt. The article references OpenAI’s multi‑level monitoring system as a possible model, where each token generated by an agent passes through a that can escalate suspicious content to more powerful AI reviewers. Anthropic’s exact implementation remains undisclosed.
Why it matters
Shifting supervision from humans to an AI model changes the safety architecture of Claude, raising both efficiency gains and new risk vectors. By removing the human “Allow?” checkpoint, Anthropic hopes to handle the volume of AI‑generated code and research that exceeds human review capacity. However, the move also concentrates decision‑making in an opaque system, making it harder for users to intervene when the model behaves unexpectedly. The change could accelerate enterprise adoption of Claude for continuous workflows, but regulators and security researchers may scrutinize the lack of transparent oversight, especially after prior reports of monitoring backdoors in similar agents.
The removal of the human approval step represents a fundamental shift in Claude’s safety architecture. While it may allow Anthropic to scale Claude’s use in high‑throughput environments such as continuous integration pipelines or research assistants, it also reduces the immediate human check that can stop unintended actions.
Consolidating safety decisions in an opaque AI system raises concerns about accountability and transparency. If the AI supervisor makes an erroneous judgment, users may have limited recourse to intervene, potentially leading to the execution of harmful code or data‑exfiltration.
The change could influence enterprise procurement decisions. Companies that require strict human oversight for compliance may hesitate to adopt Claude, while those prioritizing speed and automation might find the new mode attractive.
Regulatory bodies have begun focusing on AI systems that act autonomously without explicit human consent. Anthropic’s move may trigger scrutiny under emerging frameworks that demand and human‑in‑the‑loop safeguards.
Interactive Mechanism: How It Actually Works
Explore the underlying technology behind this development interactively.
crm_get_transaction(id='4092').Which component of an AI application is the machine-learning model itself?
What to watch next
Watch for Anthropic’s rollout details, including any real‑time monitoring or interval‑based safety checks it implements. Follow user feedback on false‑positive or false‑negative interventions, and any incidents where the autonomous mode executes harmful actions. Regulators may issue guidance on AI‑driven execution without explicit human consent, and competitors could respond with alternative supervision models.
Whether Anthropic adds interval‑based safety checks or other monitoring layers to compensate for the loss of the "Allow?" prompt.
Reports of any incidents where the autonomous mode executes unsafe or unintended actions, especially in high‑risk domains like code generation or data handling.
Feedback from early adopters regarding false‑positive blocks that could hinder productivity, or false‑negative failures that could expose users to risk.
Potential regulatory responses or industry standards that address AI‑driven execution without explicit human approval.