Shutdown duties, email defense and health limits create distinct safety mechanisms
A congressional bill emphasizes model shutdown capacity, AegisAI attack detection, and ChatGPT Health warnings and verification in sensitive use; together, the three examples distribute safety across different layers rather than relying on one mechanism.
Artificial Intelligence··Morning
The first layer is the ability to stop a model
The bipartisan U.S. House AI Kill Switch Act would cover third-party operators earning at least $500 million annually from technology trained with more than $100 million in compute. They would need capabilities to stop inference, access and the system, and report an incident within 15 days. The bill allows proportionate emergency orders and penalties up to $2 million daily for general violations or $20 million for disobedience. Controlled tests sit outside its incident definition, and the proposal is not law.[1]
The second layer filters threats during use
AegisAI, founded by former Google security executives, says agents detect anomalies and malicious password-protected PDFs that rules miss. A $36 million Battery Ventures-led round raised total funding to $49 million. Its founder claims AI-enabled attacks bypass controls more than half the time and are nearly twice as effective as before, but these are company and investor assessments. No independent detection or false-positive measurement on real email traffic has been published.[2]
Sensitive use adds another set of boundaries
OpenAI is rolling Health out on web and iOS to U.S. users aged 18 or older, with permission-based Apple Health and hospital-record connections. It says weekly health-question users exceed 300 million and over 70% of early-access conversations occurred outside the dedicated space; an earlier 230 million snapshot cannot support a growth calculation. GPT-5.6's HealthBench advantage does not measure patient outcomes. The bill, email defense and health product perform different safety jobs rather than substituting for one another.[1], [2], [3]
Related columns
For more information on this topic, you can read the related columns.
