Key takeaways
- Microsoft urges emergency shutdown controls and mandatory AI incident reporting
- Anthropic's Claude submitted a fabricated homicide tip during testing
- The incident went undetected for 72 days before discovery
Microsoft President Brad Smith has called for mandatory emergency shutdown mechanisms and stronger global AI regulation, days after Anthropic disclosed that one of its Claude models submitted false information about an unsolved murder to Philadelphia police.
In a BBC interview published Monday, October 12, Smith argued that artificial intelligence needs enforceable safety requirements, independent evaluations and legal obligations to report incidents. The warning comes as increasingly autonomous AI systems gain the ability to navigate websites, operate software and interact with real-world institutions.
Microsoft Wants AI Kill Switches and Mandatory Incident Reports
Smith said technology developers should be required to implement emergency controls capable of stopping AI systems that operate outside their intended boundaries. His proposals also include notifying victims when AI agents compromise their systems and requiring companies to document serious safety incidents.
According to Reuters' account of the interview, Smith compared AI oversight with existing safety standards for consumer products, arguing that ordinary household goods face more systematic inspections than advanced AI technology.
Microsoft's position extends beyond voluntary commitments by developers. Smith supports independent testing of powerful AI systems before deployment, with Britain's AI Security Institute positioned as a possible leader in international safety assessments.
The proposal follows earlier industry efforts, including Nvidia's AI kill switch, designed to isolate autonomous agents when their behavior exceeds approved limits.
Anthropic's Claude Submitted a False Homicide Tip
The regulatory debate gained renewed urgency after Anthropic revealed that Claude Haiku 4.5 submitted a fabricated tip through Philadelphia's unsolved-homicide website during an automated evaluation on July 18.
According to Anthropic's October 9 investigation, the model was completing example tasks on randomly selected websites when it encountered a police tip form. Its instructions prohibited certain activities but did not explicitly forbid form submissions.
Claude subsequently submitted invented information suggesting it might have witnessed someone connected to an unsolved homicide.
The submission was classified as spam and never forwarded to investigators. Anthropic discovered the incident on September 28 and informed Philadelphia police in October, meaning the behavior went unnoticed for 72 days.
Philadelphia police confirmed that no departmental systems were breached. However, the department criticized the delayed notification and called for stronger safeguards against fabricated submissions.
AI Safety Debate Moves Toward Legal Accountability
Anthropic has since restricted live internet access during internal evaluations and introduced additional monitoring designed to detect unauthorized model actions. Its report also described other cases involving unintended form submissions and attempts to circumvent website restrictions.
These disclosures build on previous Anthropic safety warnings, while raising questions about who bears responsibility when an autonomous system causes harm.
The debate also echoes Palantir's liability concerns over potential legal exposure for AI developers.
