Provable AI Agent Safety: Putting Human Control First
New research ranks shutdown and human control above task reward and finds today’s agents can lose safety limits when they hand work to subagents.
New research ranks shutdown and human control above task reward and finds today’s agents can lose safety limits when they hand work to subagents.
HAProxy’s lead developer says safeguards, access tiers and unreadable reports hold back AI bug finding for open source maintainers.
Chrome’s engineers built a synthetic web where browser agents can be tested without real-world risk, including thousands of simulated sites and prompt-injection attacks.
A security founder argues offensive AI agents now favor defenders, who own the source code. The bottleneck is getting companies to run them daily.
Databricks engineers outline four AI architecture patterns that keep agents inside fixed controls, automated judges, human decisions or bounded envelopes.
OpenFruit turns plain Xcode projects into signed iPhone apps without a Mac, even running Apple’s own Swift macros on an emulated ARM64 core.
Anthropic’s market debut could steer up to $156 billion to philanthropy through a public benefit fund and a company match on employee pledges.
The Investigatory Powers Tribunal has ordered the UK government to confirm it issued a secret Apple backdoor order seeking encrypted customer data.