EverydayApparatus
ReadsListenConceptsBrief
AllIntelligenceSoftwareSecurityPhysicsMathematicsLifeChemistrySociety

Security

1 month agoarXivSecurity

The AI Security Scanner That Learned to Sound Sure without Learning Anything

Researchers trained AI to catch dangerous bugs in the software that runs the internet — and found it performs barely better than a coin flip, for a reason that should worry anyone building these tools.
1 month agoarXivSecurity

The Best Defense against AI Hackers Is to Lie to Them

A new security strategy flips the script: instead of refusing suspicious requests, fool the attacker's AI judge into thinking it already won.
1 month agoarXivSecurity

The Math Baked into AI Safety Monitors Was Making Risk Look Smaller than It Is

When we built safety systems for AI agents, we quietly assumed the worst could never happen all at once. We were wrong, and now there's a fix.
1 month agoarXivSecurity

We're Handing AI Agents Master Keys. That Should Terrify Us.

Security built for human employees — check the ID at the door, trust what's inside — was never meant for agents that never stop acting.