Useful agents. Tested trust.

Rethinking Agents is an independent publication investigating how useful agents fail—and which defenses preserve their usefulness.

We write for engineers connecting agents to tools, code and private data, and the security reviewers helping them decide what to enable.

Security in working systems

Our starting question is practical: what can go wrong when this agent gets access, and what should change before we trust it? Reliability, architecture, cost and human oversight belong here when they change that decision.

Successful designs deserve scrutiny alongside failures. A defense that prevents the useful task needs an explicit tradeoff, not a victory lap.

How we work

We follow the useful task, available authority, failure mechanism and proposed intervention. Then we ask what useful work remains and what evidence is still missing.

• Link primary sources and record the dates and versions that matter.
• Attribute reported discoveries to their researchers. Our architectural analysis is identified separately.
• Label illustrative scenarios and proposed controls. Do not imply they were tested.
• For original experiments, report setup, baselines, attempts, outcomes and limitations.
• Correct consequential errors and explain material updates.

What you can read here

The opening collection combines documented cases with conditional design analysis. It contains no claimed original attack reproduction or defense benchmark.

Read the research collection, or send a source, correction or practical question to hello@rethinkingagents.com.